{"font_size":0.4,"font_color":"#FFFFFF","background_alpha":0.5,"background_color":"#9C27B0","Stroke":"none","body":[{"from":4.72,"to":8.01,"location":2,"content":"okay hi everyone let's get started so"},{"from":8.01,"to":10.36,"location":2,"content":"Chris is traveling this week so he's not"},{"from":10.36,"to":12.43,"location":2,"content":"here but I'm very excited to say that"},{"from":12.43,"to":14.56,"location":2,"content":"today we've got Margaret Mitchell who is"},{"from":14.56,"to":16.96,"location":2,"content":"a senior research scientist at Google AI"},{"from":16.96,"to":19.72,"location":2,"content":"and she's gonna tell us about the latest"},{"from":19.72,"to":22.57,"location":2,"content":"work defining and understanding and"},{"from":22.57,"to":25.18,"location":2,"content":"improving the situation with bias in"},{"from":25.18,"to":28.12,"location":2,"content":"artificial intelligence Margaret has a"},{"from":28.12,"to":29.86,"location":2,"content":"background working in NLP and deep"},{"from":29.86,"to":31.3,"location":2,"content":"learning so I'm really interested to"},{"from":31.3,"to":33.07,"location":2,"content":"hear what she has to say today take it"},{"from":33.07,"to":33.4,"location":2,"content":"away"},{"from":33.4,"to":36.04,"location":2,"content":"great thank you and can you guys hear me"},{"from":36.04,"to":38.14,"location":2,"content":"okay I'm not sure if this mic is exactly"},{"from":38.14,"to":40.66,"location":2,"content":"picking up my bus everything's cool okay"},{"from":40.66,"to":45.34,"location":2,"content":"cool um so this work is the product of a"},{"from":45.34,"to":46.24,"location":2,"content":"ton of different people and"},{"from":46.24,"to":47.74,"location":2,"content":"collaborators that I've tried to put up"},{"from":47.74,"to":50.62,"location":2,"content":"here some students at Stanford also"},{"from":50.62,"to":54.3,"location":2,"content":"Johns Hopkins Google Facebook and"},{"from":54.3,"to":61.17,"location":2,"content":"Microsoft are all represented cool so um"},{"from":61.17,"to":64.06,"location":2,"content":"for those of you who haven't seen this"},{"from":64.06,"to":67.15,"location":2,"content":"set of slides before what do you see"},{"from":67.15,"to":71.26,"location":2,"content":"here just shout it up bananas okay what"},{"from":71.26,"to":72.81,"location":2,"content":"else"},{"from":72.81,"to":94.12,"location":2,"content":"stickers what else bananas with stickers"},{"from":94.12,"to":95.59,"location":2,"content":"on them you can start doing like"},{"from":95.59,"to":97.81,"location":2,"content":"embedded clauses you know bunches of"},{"from":97.81,"to":99.46,"location":2,"content":"bananas with stickers on them on shelves"},{"from":99.46,"to":101.71,"location":2,"content":"in the store to get kind of crazy but we"},{"from":101.71,"to":104.35,"location":2,"content":"don't tend to say yellow bananas right"},{"from":104.35,"to":107.26,"location":2,"content":"so give them something like this we"},{"from":107.26,"to":109.63,"location":2,"content":"might say green bananas or we might say"},{"from":109.63,"to":112.9,"location":2,"content":"unripe bananas given an image like this"},{"from":112.9,"to":115.66,"location":2,"content":"we might say ripe bananas or bananas"},{"from":115.66,"to":118.69,"location":2,"content":"with spots on them if you're me you"},{"from":118.69,"to":119.74,"location":2,"content":"might say bananas that are good for"},{"from":119.74,"to":122.98,"location":2,"content":"banana bread and but given an image like"},{"from":122.98,"to":124.66,"location":2,"content":"this or something like this in the real"},{"from":124.66,"to":126.82,"location":2,"content":"world we tend not to mention the"},{"from":126.82,"to":129.13,"location":2,"content":"yellowness and the reason for this is"},{"from":129.13,"to":131.62,"location":2,"content":"because yellow is prototypical for"},{"from":131.62,"to":135.82,"location":2,"content":"bananas so the idea of prototypes stems"},{"from":135.82,"to":137.62,"location":2,"content":"from prototype theory which goes back to"},{"from":137.62,"to":138.46,"location":2,"content":"the early"},{"from":138.46,"to":140.74,"location":2,"content":"coming out of the work of Eleanor Rosch"},{"from":140.74,"to":143.44,"location":2,"content":"and colleagues and it's this idea that"},{"from":143.44,"to":145.83,"location":2,"content":"there are some stored central"},{"from":145.83,"to":149.5,"location":2,"content":"prototypical notions of objects that we"},{"from":149.5,"to":152.35,"location":2,"content":"access as we're operating throughout the"},{"from":152.35,"to":154.6,"location":2,"content":"world there's some disagreement about"},{"from":154.6,"to":157.65,"location":2,"content":"whether these prototypes are actual"},{"from":157.65,"to":160.09,"location":2,"content":"exemplars of objects or something like a"},{"from":160.09,"to":162.52,"location":2,"content":"distribution over what's likely but"},{"from":162.52,"to":164.29,"location":2,"content":"there is general agreement that we do"},{"from":164.29,"to":166.21,"location":2,"content":"have some sort of sense of what's"},{"from":166.21,"to":168.91,"location":2,"content":"typical and what's a typical of the"},{"from":168.91,"to":170.8,"location":2,"content":"things in the world and we tend to"},{"from":170.8,"to":172.66,"location":2,"content":"notice and talk about the things that"},{"from":172.66,"to":179.2,"location":2,"content":"are atypical so this is a riddle that I"},{"from":179.2,"to":181,"location":2,"content":"heard in middle school that worked a"},{"from":181,"to":183.46,"location":2,"content":"little bit more at that time some of you"},{"from":183.46,"to":185.68,"location":2,"content":"might have heard it before a man and his"},{"from":185.68,"to":187.72,"location":2,"content":"son are in a terrible accident and are"},{"from":187.72,"to":189.4,"location":2,"content":"rushed to the hospital in critical care"},{"from":189.4,"to":191.98,"location":2,"content":"the doctor looks at the boy and exclaims"},{"from":191.98,"to":194.44,"location":2,"content":"I can't operate on this boy he's my son"},{"from":194.44,"to":200.68,"location":2,"content":"how could this be - Deb's or more he has"},{"from":200.68,"to":203.65,"location":2,"content":"them on her doctor right otherwise known"},{"from":203.65,"to":206.53,"location":2,"content":"as a female doctor which might be"},{"from":206.53,"to":211.09,"location":2,"content":"contract is contrasted with doctor in a"},{"from":211.09,"to":213.07,"location":2,"content":"study they did when they first sort of"},{"from":213.07,"to":215.62,"location":2,"content":"put forward this riddle at Boston"},{"from":215.62,"to":217.18,"location":2,"content":"University they found that the majority"},{"from":217.18,"to":219.1,"location":2,"content":"of test subjects overlooked the"},{"from":219.1,"to":220.87,"location":2,"content":"possibility that the doctor could be a"},{"from":220.87,"to":223.66,"location":2,"content":"she and that included men women and"},{"from":223.66,"to":226.48,"location":2,"content":"self-described feminists so the point is"},{"from":226.48,"to":230.35,"location":2,"content":"that these kinds of ways of talking"},{"from":230.35,"to":232,"location":2,"content":"about things and assumptions that we"},{"from":232,"to":234.67,"location":2,"content":"make aren't necessarily something that"},{"from":234.67,"to":238.6,"location":2,"content":"speaks to negative intent but something"},{"from":238.6,"to":240.48,"location":2,"content":"that speaks to how we actually store"},{"from":240.48,"to":242.79,"location":2,"content":"representations in our minds and how we"},{"from":242.79,"to":245.08,"location":2,"content":"access those representations as we"},{"from":245.08,"to":248.92,"location":2,"content":"interact in the world so this this"},{"from":248.92,"to":250.69,"location":2,"content":"affects what we can learn when we're"},{"from":250.69,"to":254.29,"location":2,"content":"learning from text so this is work from"},{"from":254.29,"to":257.05,"location":2,"content":"2013 where they took a look at what was"},{"from":257.05,"to":258.73,"location":2,"content":"sort of most likely what would you learn"},{"from":258.73,"to":261.3,"location":2,"content":"if you were just learning from raw text"},{"from":261.3,"to":263.47,"location":2,"content":"what were some things that were common"},{"from":263.47,"to":267.34,"location":2,"content":"in the world um and they found that in"},{"from":267.34,"to":269.62,"location":2,"content":"this set up something like murdering was"},{"from":269.62,"to":270.86,"location":2,"content":"ten times"},{"from":270.86,"to":273.35,"location":2,"content":"likely than blinking and the reason for"},{"from":273.35,"to":275.06,"location":2,"content":"this is because people tend not to"},{"from":275.06,"to":277.34,"location":2,"content":"mention these typical things that go"},{"from":277.34,"to":280.07,"location":2,"content":"without saying we don't tend to mention"},{"from":280.07,"to":282.77,"location":2,"content":"things like blinking and breathing but"},{"from":282.77,"to":284.96,"location":2,"content":"we do mention atypical events like"},{"from":284.96,"to":287.21,"location":2,"content":"murder and that affects the kind of"},{"from":287.21,"to":289.46,"location":2,"content":"things a machine can learn from text"},{"from":289.46,"to":291.23,"location":2,"content":"that we put out in the world because"},{"from":291.23,"to":292.61,"location":2,"content":"it's been subject to all of these"},{"from":292.61,"to":294.44,"location":2,"content":"filtering processes that we have as"},{"from":294.44,"to":298.82,"location":2,"content":"humans before we communicate this issue"},{"from":298.82,"to":300.71,"location":2,"content":"in particular is known as human"},{"from":300.71,"to":302.36,"location":2,"content":"reporting bias which is that the"},{"from":302.36,"to":304.25,"location":2,"content":"frequency with which people write about"},{"from":304.25,"to":307.07,"location":2,"content":"actions outcomes or properties is not a"},{"from":307.07,"to":309.26,"location":2,"content":"reflection of real-world frequencies or"},{"from":309.26,"to":310.82,"location":2,"content":"the degree to which a property is"},{"from":310.82,"to":312.41,"location":2,"content":"characteristic of a class of individuals"},{"from":312.41,"to":314.6,"location":2,"content":"but says a lot more about how we're"},{"from":314.6,"to":316.22,"location":2,"content":"actually processing the world and what"},{"from":316.22,"to":320.06,"location":2,"content":"we think is remarkable so this affects"},{"from":320.06,"to":322.88,"location":2,"content":"everything a system can learn in a"},{"from":322.88,"to":324.89,"location":2,"content":"typical machine learning paradigm one of"},{"from":324.89,"to":326.48,"location":2,"content":"the first steps is to collect and"},{"from":326.48,"to":330.23,"location":2,"content":"potentially annotate training data from"},{"from":330.23,"to":334.73,"location":2,"content":"there a model can be trained from there"},{"from":334.73,"to":337.49,"location":2,"content":"media can be filtered ranked ranked"},{"from":337.49,"to":340.52,"location":2,"content":"aggregated generated in some way and"},{"from":340.52,"to":342.98,"location":2,"content":"from there people see the output and we"},{"from":342.98,"to":345.11,"location":2,"content":"like to think of this as a relatively"},{"from":345.11,"to":347.66,"location":2,"content":"straightforward pipeline but at the very"},{"from":347.66,"to":350.6,"location":2,"content":"start even before we're collecting with"},{"from":350.6,"to":352.49,"location":2,"content":"the data actually within the data itself"},{"from":352.49,"to":355.07,"location":2,"content":"are a host of different kinds of human"},{"from":355.07,"to":357.53,"location":2,"content":"biases so things like stereotyping"},{"from":357.53,"to":359,"location":2,"content":"things like prejudice things like a"},{"from":359,"to":361.34,"location":2,"content":"racism and that's embedded within the"},{"from":361.34,"to":363.92,"location":2,"content":"data before we collect it then as we"},{"from":363.92,"to":366.53,"location":2,"content":"collect and annotate data further biases"},{"from":366.53,"to":368.45,"location":2,"content":"become introduced so things like"},{"from":368.45,"to":371.68,"location":2,"content":"sampling errors confirmation bias"},{"from":371.68,"to":374.12,"location":2,"content":"in-group bias and out-group bias and"},{"from":374.12,"to":376.04,"location":2,"content":"I'll talk about these a little bit oh"},{"from":376.04,"to":378.02,"location":2,"content":"and I should mention feel free to ask"},{"from":378.02,"to":381.26,"location":2,"content":"questions as I go totally fine to just"},{"from":381.26,"to":385.34,"location":2,"content":"kind of interact throughout so here are"},{"from":385.34,"to":386.89,"location":2,"content":"some of the biases that I think are"},{"from":386.89,"to":389.48,"location":2,"content":"relatively important for work in AI"},{"from":389.48,"to":391.58,"location":2,"content":"machine learning there's hundreds you"},{"from":391.58,"to":393.89,"location":2,"content":"can go into but some of the ones that"},{"from":393.89,"to":395.51,"location":2,"content":"I've sort of become the most aware of"},{"from":395.51,"to":398,"location":2,"content":"working in this space are this set and"},{"from":398,"to":400.4,"location":2,"content":"I'll go through each of these a bit so I"},{"from":400.4,"to":401.99,"location":2,"content":"talked about reporting bias earlier"},{"from":401.99,"to":404.18,"location":2,"content":"which is which affects what we can learn"},{"from":404.18,"to":406.74,"location":2,"content":"from"},{"from":406.74,"to":409.36,"location":2,"content":"another example of a kind of bias that"},{"from":409.36,"to":410.59,"location":2,"content":"really affects what we can learn from"},{"from":410.59,"to":414.22,"location":2,"content":"text is selection bias so a lot of times"},{"from":414.22,"to":415.96,"location":2,"content":"that we a lot of times when we get data"},{"from":415.96,"to":418.57,"location":2,"content":"annotated we do something like Amazon's"},{"from":418.57,"to":421.06,"location":2,"content":"Mechanical Turk and the distribution of"},{"from":421.06,"to":423.25,"location":2,"content":"workers across the world is not an even"},{"from":423.25,"to":425.47,"location":2,"content":"sort of uniform distribution it's"},{"from":425.47,"to":428.65,"location":2,"content":"actually concentrated in India the US"},{"from":428.65,"to":430.45,"location":2,"content":"and then some in Europe so this leaves"},{"from":430.45,"to":432.94,"location":2,"content":"out South America this leaves out Africa"},{"from":432.94,"to":435.25,"location":2,"content":"this leaves out a lot of China and that"},{"from":435.25,"to":436.75,"location":2,"content":"affects the kind of things that we'll be"},{"from":436.75,"to":438.34,"location":2,"content":"able to learn about the world when we"},{"from":438.34,"to":443.41,"location":2,"content":"have things annotated another kind of"},{"from":443.41,"to":445.39,"location":2,"content":"bias is out-group homogeneity bias which"},{"from":445.39,"to":447.19,"location":2,"content":"is the tendency to see out group members"},{"from":447.19,"to":449.32,"location":2,"content":"as more alike than in-group members and"},{"from":449.32,"to":451.39,"location":2,"content":"this is going to affect what people are"},{"from":451.39,"to":453.28,"location":2,"content":"able to describe and talk about when"},{"from":453.28,"to":454.63,"location":2,"content":"they're annotating things such as"},{"from":454.63,"to":457.72,"location":2,"content":"emotion so so for example we have these"},{"from":457.72,"to":460.3,"location":2,"content":"two like adorable puppies on the left"},{"from":460.3,"to":461.62,"location":2,"content":"here and they're looking at these four"},{"from":461.62,"to":464.23,"location":2,"content":"cats and these are all different black"},{"from":464.23,"to":465.73,"location":2,"content":"cats very different in different ways"},{"from":465.73,"to":468.07,"location":2,"content":"but the two puppies look at the cats and"},{"from":468.07,"to":470.44,"location":2,"content":"they see four cats basically the same"},{"from":470.44,"to":472.51,"location":2,"content":"and it's kind of trivial to understand"},{"from":472.51,"to":474.34,"location":2,"content":"how that also extends to human cognition"},{"from":474.34,"to":477.79,"location":2,"content":"and how we also process people it's this"},{"from":477.79,"to":480.15,"location":2,"content":"it's the sense we have that the the"},{"from":480.15,"to":482.56,"location":2,"content":"cohort that we're in the people that we"},{"from":482.56,"to":484.54,"location":2,"content":"interact with those are the kinds of"},{"from":484.54,"to":486.73,"location":2,"content":"people that are nuanced and everybody"},{"from":486.73,"to":489.43,"location":2,"content":"else is somehow less nuanced has less"},{"from":489.43,"to":491.71,"location":2,"content":"detail to them it's a trick our minds"},{"from":491.71,"to":493.72,"location":2,"content":"play on us in order to help us process"},{"from":493.72,"to":495.94,"location":2,"content":"the world but it affects how we talk"},{"from":495.94,"to":497.53,"location":2,"content":"about it and it affects further how we"},{"from":497.53,"to":502.18,"location":2,"content":"annotate it this leads to stuff like"},{"from":502.18,"to":504.55,"location":2,"content":"bias data representations so it's"},{"from":504.55,"to":506.08,"location":2,"content":"possible that you have an appropriate"},{"from":506.08,"to":508.63,"location":2,"content":"amount of data for every possible human"},{"from":508.63,"to":511.42,"location":2,"content":"group you can think of in your data but"},{"from":511.42,"to":512.86,"location":2,"content":"it might be the case that some groups"},{"from":512.86,"to":514.54,"location":2,"content":"are represented less positively than"},{"from":514.54,"to":516.49,"location":2,"content":"others and if we have time I'll go into"},{"from":516.49,"to":521.29,"location":2,"content":"a long a longer example of that it also"},{"from":521.29,"to":523.47,"location":2,"content":"leads to things like biased labels so"},{"from":523.47,"to":526.06,"location":2,"content":"this is a issue that came up when we"},{"from":526.06,"to":527.62,"location":2,"content":"were getting some annotations for"},{"from":527.62,"to":530.14,"location":2,"content":"inclusive images competition asking"},{"from":530.14,"to":532.57,"location":2,"content":"people to annotate things like bride and"},{"from":532.57,"to":535.45,"location":2,"content":"wedding and groom and we found that"},{"from":535.45,"to":537.01,"location":2,"content":"given three different kinds of bride"},{"from":537.01,"to":537.48,"location":2,"content":"wedding"},{"from":537.48,"to":541.17,"location":2,"content":"room images ones that were more Western"},{"from":541.17,"to":544.95,"location":2,"content":"European American got the appropriate"},{"from":544.95,"to":547.38,"location":2,"content":"labels and ones that weren't just got"},{"from":547.38,"to":549.63,"location":2,"content":"sort of more generic person kinds of"},{"from":549.63,"to":552.75,"location":2,"content":"labels not able to actually tease out"},{"from":552.75,"to":554.25,"location":2,"content":"what's actually happening in these"},{"from":554.25,"to":560.25,"location":2,"content":"images companies issue our biases in"},{"from":560.25,"to":562.59,"location":2,"content":"interpretation when the model outputs"},{"from":562.59,"to":566.55,"location":2,"content":"its decisions so one one issue is"},{"from":566.55,"to":568.44,"location":2,"content":"confirmation bias which is the tendency"},{"from":568.44,"to":571.17,"location":2,"content":"to search for interpret favor recall"},{"from":571.17,"to":572.67,"location":2,"content":"information in a way that confirms"},{"from":572.67,"to":574.83,"location":2,"content":"pre-existing beliefs and so a lot of"},{"from":574.83,"to":577.86,"location":2,"content":"times when we build and to end systems"},{"from":577.86,"to":580.32,"location":2,"content":"and try and test our hypotheses were"},{"from":580.32,"to":582.54,"location":2,"content":"kind of just testing it towards things"},{"from":582.54,"to":584.79,"location":2,"content":"that we want to be true and analyzing"},{"from":584.79,"to":586.44,"location":2,"content":"the results in a way that will help"},{"from":586.44,"to":590.34,"location":2,"content":"confirm what we want to be true over"},{"from":590.34,"to":592.14,"location":2,"content":"generalization which is coming to a"},{"from":592.14,"to":593.67,"location":2,"content":"conclusion based on information that's"},{"from":593.67,"to":596.37,"location":2,"content":"too general or not specific enough this"},{"from":596.37,"to":597.96,"location":2,"content":"is an issue that happens a lot of times"},{"from":597.96,"to":600.69,"location":2,"content":"in the analysis of deep learning model"},{"from":600.69,"to":603.51,"location":2,"content":"results where it's assumed that there's"},{"from":603.51,"to":605.94,"location":2,"content":"there's some kind of general conclusion"},{"from":605.94,"to":607.62,"location":2,"content":"that can be taken away when really it's"},{"from":607.62,"to":609.84,"location":2,"content":"actually just an effect of really skewed"},{"from":609.84,"to":612.57,"location":2,"content":"data this is also closely related to"},{"from":612.57,"to":614.73,"location":2,"content":"overfitting which is kind of the machine"},{"from":614.73,"to":616.59,"location":2,"content":"learning version of over generalization"},{"from":616.59,"to":618.72,"location":2,"content":"which is where you're still making"},{"from":618.72,"to":620.55,"location":2,"content":"predictions and outcomes but it's based"},{"from":620.55,"to":624.09,"location":2,"content":"on a small set of possible features so"},{"from":624.09,"to":626.34,"location":2,"content":"it's not actually capturing the space of"},{"from":626.34,"to":629.22,"location":2,"content":"the correct features for the outcome the"},{"from":629.22,"to":632.75,"location":2,"content":"desired output prediction correctly"},{"from":632.75,"to":635.46,"location":2,"content":"there's also correlation fallacy which"},{"from":635.46,"to":637.14,"location":2,"content":"is confusing correlation with causation"},{"from":637.14,"to":639.75,"location":2,"content":"and this happens a lot again in talking"},{"from":639.75,"to":641.43,"location":2,"content":"about what machine learning models are"},{"from":641.43,"to":642.9,"location":2,"content":"learning and deep learning models are"},{"from":642.9,"to":645,"location":2,"content":"learning in particular where just"},{"from":645,"to":647.46,"location":2,"content":"because things happen together doesn't"},{"from":647.46,"to":649.41,"location":2,"content":"mean that one is causing the other but"},{"from":649.41,"to":651.72,"location":2,"content":"models don't tell you anything deep"},{"from":651.72,"to":653.13,"location":2,"content":"learning models directly don't tell you"},{"from":653.13,"to":655.11,"location":2,"content":"anything about the causal relations and"},{"from":655.11,"to":657.15,"location":2,"content":"so it's easy to think that some output"},{"from":657.15,"to":659.1,"location":2,"content":"that is predicted based on a correlation"},{"from":659.1,"to":661.08,"location":2,"content":"is actually something that's causal and"},{"from":661.08,"to":662.91,"location":2,"content":"I'll talk about some examples of this"},{"from":662.91,"to":663.39,"location":2,"content":"too"},{"from":663.39,"to":667.23,"location":2,"content":"a further issue is automation bias and"},{"from":667.23,"to":669.15,"location":2,"content":"this really affects the machine learning"},{"from":669.15,"to":671.1,"location":2,"content":"models we put out there in the world"},{"from":671.1,"to":673.59,"location":2,"content":"get used by people in systems like"},{"from":673.59,"to":676.38,"location":2,"content":"justice systems so that's the tendency"},{"from":676.38,"to":680.4,"location":2,"content":"to favor the suggestions of automatic"},{"from":680.4,"to":682.41,"location":2,"content":"predictions of models that output"},{"from":682.41,"to":686.94,"location":2,"content":"predictions over the over the different"},{"from":686.94,"to":689.31,"location":2,"content":"kinds of suggestions of another human"},{"from":689.31,"to":691.47,"location":2,"content":"and this happens even in the face of"},{"from":691.47,"to":694.2,"location":2,"content":"contradictory evidence so if a system is"},{"from":694.2,"to":696.66,"location":2,"content":"telling you you know this this is the"},{"from":696.66,"to":699.81,"location":2,"content":"score or this is the risk of this"},{"from":699.81,"to":701.76,"location":2,"content":"individual then we're more likely to"},{"from":701.76,"to":703.95,"location":2,"content":"think it's true because it came out of a"},{"from":703.95,"to":706.86,"location":2,"content":"mathematical system and we automatically"},{"from":706.86,"to":708.15,"location":2,"content":"sort of see this as something more"},{"from":708.15,"to":709.98,"location":2,"content":"objective something more mathematical"},{"from":709.98,"to":711.87,"location":2,"content":"that something's going to be more true"},{"from":711.87,"to":714.42,"location":2,"content":"than human some somehow and that's"},{"from":714.42,"to":718.41,"location":2,"content":"automation bias so rather than this kind"},{"from":718.41,"to":720.93,"location":2,"content":"of clean straightforward pipeline that"},{"from":720.93,"to":723.03,"location":2,"content":"we have in machine learning and we have"},{"from":723.03,"to":724.68,"location":2,"content":"human bias coming in at the very start"},{"from":724.68,"to":728.43,"location":2,"content":"in the data and then human bias coming"},{"from":728.43,"to":730.5,"location":2,"content":"in in data collection annotation and"},{"from":730.5,"to":732.3,"location":2,"content":"then further getting propagated through"},{"from":732.3,"to":734.85,"location":2,"content":"the system as we train on that data as"},{"from":734.85,"to":737.01,"location":2,"content":"we start putting outputs based on that"},{"from":737.01,"to":739.5,"location":2,"content":"data as people act on that data and this"},{"from":739.5,"to":742.62,"location":2,"content":"creates a feedback loop where the kinds"},{"from":742.62,"to":744.57,"location":2,"content":"of things that we output for people to"},{"from":744.57,"to":748.29,"location":2,"content":"act on are then are then then serves as"},{"from":748.29,"to":750.69,"location":2,"content":"further training data for input for new"},{"from":750.69,"to":753.12,"location":2,"content":"system so you end up amplifying even"},{"from":753.12,"to":754.71,"location":2,"content":"further these different kinds of"},{"from":754.71,"to":758.16,"location":2,"content":"implicit biases this is known as a bias"},{"from":758.16,"to":760.71,"location":2,"content":"Network effect or bias Laundering I like"},{"from":760.71,"to":765.12,"location":2,"content":"to call it and so the message is that"},{"from":765.12,"to":767.49,"location":2,"content":"human data perpetuates human biases and"},{"from":767.49,"to":769.38,"location":2,"content":"then as as machine learning or deep"},{"from":769.38,"to":771.48,"location":2,"content":"learning learns from human data the"},{"from":771.48,"to":775.08,"location":2,"content":"result is a bias network effect so I"},{"from":775.08,"to":777.51,"location":2,"content":"want to steer clear the idea that if I"},{"from":777.51,"to":780.06,"location":2,"content":"say bias or if someone says bias that"},{"from":780.06,"to":782.01,"location":2,"content":"equals bad it's a little bit more"},{"from":782.01,"to":784.68,"location":2,"content":"nuanced than that so there are all kinds"},{"from":784.68,"to":786.42,"location":2,"content":"of things that people mean when they're"},{"from":786.42,"to":788.91,"location":2,"content":"talking about bias and even the same"},{"from":788.91,"to":790.95,"location":2,"content":"bias can be good in some situations and"},{"from":790.95,"to":793.38,"location":2,"content":"bad in some situations so bias in"},{"from":793.38,"to":795.66,"location":2,"content":"statistics on ml we could we talked"},{"from":795.66,"to":797.37,"location":2,"content":"about the bias of an estimator which is"},{"from":797.37,"to":798.78,"location":2,"content":"the difference between the predictions"},{"from":798.78,"to":800.67,"location":2,"content":"and the and the truth for ground truth"},{"from":800.67,"to":803.32,"location":2,"content":"we talked about the bias term in linear"},{"from":803.32,"to":806.71,"location":2,"content":"rushon we also have cognitive biases and"},{"from":806.71,"to":808.27,"location":2,"content":"I talked about that in the beginning and"},{"from":808.27,"to":810.61,"location":2,"content":"not all of those are negative or or have"},{"from":810.61,"to":813.52,"location":2,"content":"to be or have to be seen as negative so"},{"from":813.52,"to":816.16,"location":2,"content":"optimism is another kind of bias that we"},{"from":816.16,"to":818.05,"location":2,"content":"can have that affects our worldview and"},{"from":818.05,"to":820.18,"location":2,"content":"the way we sort of process things and"},{"from":820.18,"to":822.01,"location":2,"content":"even things like recency bias and"},{"from":822.01,"to":824.29,"location":2,"content":"confirmation bias are just ways that our"},{"from":824.29,"to":827.92,"location":2,"content":"minds can like handle the combinatorial"},{"from":827.92,"to":829.6,"location":2,"content":"explosion of all the different things"},{"from":829.6,"to":831.49,"location":2,"content":"that can be true in the world and put it"},{"from":831.49,"to":832.93,"location":2,"content":"down to something tractable that we can"},{"from":832.93,"to":834.55,"location":2,"content":"sort of operate with in the real world"},{"from":834.55,"to":838.33,"location":2,"content":"and so algorithmic bias is what a lot of"},{"from":838.33,"to":840.25,"location":2,"content":"people mean and headlines and whatnot"},{"from":840.25,"to":842.17,"location":2,"content":"when they're talking about bias which is"},{"from":842.17,"to":845.11,"location":2,"content":"more about unjust unfair or prejudicial"},{"from":845.11,"to":847.24,"location":2,"content":"treatment of people that's an output of"},{"from":847.24,"to":850.24,"location":2,"content":"an automated decision system and the"},{"from":850.24,"to":853.75,"location":2,"content":"focus here is really on unjust unfair or"},{"from":853.75,"to":855.85,"location":2,"content":"prejudicial treatment of people so a lot"},{"from":855.85,"to":857.32,"location":2,"content":"of the work in this space right now is"},{"from":857.32,"to":859.57,"location":2,"content":"focusing on trying to understand what"},{"from":859.57,"to":861.67,"location":2,"content":"does it mean to be unjust from an"},{"from":861.67,"to":864.1,"location":2,"content":"algorithm what does it mean to be unfair"},{"from":864.1,"to":866.92,"location":2,"content":"from an algorithm and how can we handle"},{"from":866.92,"to":868.36,"location":2,"content":"this how can we sort of mitigate these"},{"from":868.36,"to":870.4,"location":2,"content":"issues in order to be able to keep"},{"from":870.4,"to":872.05,"location":2,"content":"developing technology that's useful for"},{"from":872.05,"to":876.78,"location":2,"content":"people without worsening social divides"},{"from":876.78,"to":879.22,"location":2,"content":"and I thought the Guardian put it really"},{"from":879.22,"to":882.16,"location":2,"content":"well a few years ago they said although"},{"from":882.16,"to":883.93,"location":2,"content":"neural networks might be said to write"},{"from":883.93,"to":886.33,"location":2,"content":"their own programs they do so towards"},{"from":886.33,"to":888.79,"location":2,"content":"goals set by humans using data collected"},{"from":888.79,"to":891.07,"location":2,"content":"for human purposes if the data is skewed"},{"from":891.07,"to":893.17,"location":2,"content":"even by accident the computers will"},{"from":893.17,"to":895.93,"location":2,"content":"amplify injustice and it really keyed in"},{"from":895.93,"to":899.02,"location":2,"content":"on this amplify and justice idea and"},{"from":899.02,"to":902.62,"location":2,"content":"let's talk about what that can mean so"},{"from":902.62,"to":904.57,"location":2,"content":"one of the avenues of deep learning"},{"from":904.57,"to":906.55,"location":2,"content":"research that's taken off in the past"},{"from":906.55,"to":908.23,"location":2,"content":"few years is predicting criminal"},{"from":908.23,"to":912.37,"location":2,"content":"behavior so how many of you are familiar"},{"from":912.37,"to":917.29,"location":2,"content":"with predictive policing okay like half"},{"from":917.29,"to":918.43,"location":2,"content":"of the class okay"},{"from":918.43,"to":921.85,"location":2,"content":"so in predictive policing algorithms are"},{"from":921.85,"to":925.09,"location":2,"content":"predict areas to deploy officers where"},{"from":925.09,"to":927.76,"location":2,"content":"crime is considered to be likely to"},{"from":927.76,"to":932.74,"location":2,"content":"occur but the data that the the models"},{"from":932.74,"to":935.41,"location":2,"content":"are trained off of is based on where"},{"from":935.41,"to":936.2,"location":2,"content":"police off"},{"from":936.2,"to":938.69,"location":2,"content":"SURS have already gone and made arrests"},{"from":938.69,"to":941.27,"location":2,"content":"so the systems are simply learning the"},{"from":941.27,"to":943.4,"location":2,"content":"patterns of bias that humans have in"},{"from":943.4,"to":945.35,"location":2,"content":"where they go and where they are trying"},{"from":945.35,"to":948.92,"location":2,"content":"to decide to defer to find crime and"},{"from":948.92,"to":951.32,"location":2,"content":"then reflecting them back so because the"},{"from":951.32,"to":953.69,"location":2,"content":"system hones in on some of the top spots"},{"from":953.69,"to":956.6,"location":2,"content":"where people have been arrested notice"},{"from":956.6,"to":958.22,"location":2,"content":"that's not the same of nuts the same"},{"from":958.22,"to":959.36,"location":2,"content":"thing as where crimes have been"},{"from":959.36,"to":961.61,"location":2,"content":"committed right it's where arrests have"},{"from":961.61,"to":964.52,"location":2,"content":"been made it means that the other areas"},{"from":964.52,"to":966.17,"location":2,"content":"that might be explored for crime don't"},{"from":966.17,"to":968,"location":2,"content":"get explored at all that worsens the"},{"from":968,"to":971.54,"location":2,"content":"situation some neighborhoods get really"},{"from":971.54,"to":974.06,"location":2,"content":"acutely focused attention on them and"},{"from":974.06,"to":975.86,"location":2,"content":"that heightens the chances of serious"},{"from":975.86,"to":978.05,"location":2,"content":"repercussions for even minor infractions"},{"from":978.05,"to":980.57,"location":2,"content":"that means arrests and that means a"},{"from":980.57,"to":982.4,"location":2,"content":"feedback loop of data that you will get"},{"from":982.4,"to":986.86,"location":2,"content":"an arrest in this place if you go there"},{"from":986.86,"to":990.65,"location":2,"content":"another sort of related issue in this"},{"from":990.65,"to":994.43,"location":2,"content":"space is predictive sentencing so there"},{"from":994.43,"to":996.02,"location":2,"content":"was a really nice article that came out"},{"from":996.02,"to":998.18,"location":2,"content":"from Pro Publica a few years ago"},{"from":998.18,"to":1001.09,"location":2,"content":"discussing this but when most defendants"},{"from":1001.09,"to":1002.68,"location":2,"content":"are booked in jail they respond to a"},{"from":1002.68,"to":1005.26,"location":2,"content":"questionnaire called compass and their"},{"from":1005.26,"to":1007.06,"location":2,"content":"answers are fed into this software"},{"from":1007.06,"to":1009.25,"location":2,"content":"system that generates scores that"},{"from":1009.25,"to":1011.2,"location":2,"content":"correspond to the risk of recidivism"},{"from":1011.2,"to":1015.09,"location":2,"content":"that's the risk of making a crime again"},{"from":1015.09,"to":1017.32,"location":2,"content":"and the questions are used to gather"},{"from":1017.32,"to":1019.26,"location":2,"content":"data on the defendants socioeconomic"},{"from":1019.26,"to":1022.42,"location":2,"content":"status family background neighborhood"},{"from":1022.42,"to":1024.55,"location":2,"content":"crime employment status and other"},{"from":1024.55,"to":1025.75,"location":2,"content":"factors in order to reach some"},{"from":1025.75,"to":1029.14,"location":2,"content":"predictive prediction of an individual's"},{"from":1029.14,"to":1033.61,"location":2,"content":"crime or criminal risk but once ends up"},{"from":1033.61,"to":1035.47,"location":2,"content":"happening is that it ends up focusing on"},{"from":1035.47,"to":1038.77,"location":2,"content":"the key bias issues that humans have and"},{"from":1038.77,"to":1041.35,"location":2,"content":"propagating it back with something that"},{"from":1041.35,"to":1044.05,"location":2,"content":"looks like an objective score so you're"},{"from":1044.05,"to":1047.17,"location":2,"content":"a lot more likely to be convicted of a"},{"from":1047.17,"to":1049.42,"location":2,"content":"crime if you're black than if you're"},{"from":1049.42,"to":1051.28,"location":2,"content":"white even if you've made the exact same"},{"from":1051.28,"to":1053.77,"location":2,"content":"crime and the system will pick up on"},{"from":1053.77,"to":1056.14,"location":2,"content":"this and will reflect this back to say"},{"from":1056.14,"to":1057.43,"location":2,"content":"that people who are black are more"},{"from":1057.43,"to":1059.38,"location":2,"content":"likely to have received like recidivism"},{"from":1059.38,"to":1061.63,"location":2,"content":"more likely to convict him to make a"},{"from":1061.63,"to":1066.34,"location":2,"content":"crime again so this is an example of"},{"from":1066.34,"to":1068.8,"location":2,"content":"automation bias preferring the output of"},{"from":1068.8,"to":1069.95,"location":2,"content":"a system"},{"from":1069.95,"to":1072.58,"location":2,"content":"in the face of overgeneralization"},{"from":1072.58,"to":1075.58,"location":2,"content":"feedback loops and correlation fallacy"},{"from":1075.58,"to":1077.57,"location":2,"content":"confusing things that are occurring"},{"from":1077.57,"to":1083.66,"location":2,"content":"together as being somehow causal there's"},{"from":1083.66,"to":1086.53,"location":2,"content":"another sort of area of research and"},{"from":1086.53,"to":1088.76,"location":2,"content":"startups looking at predicting"},{"from":1088.76,"to":1090.77,"location":2,"content":"criminality in particular some things"},{"from":1090.77,"to":1092.93,"location":2,"content":"like face images so there's a company"},{"from":1092.93,"to":1095.06,"location":2,"content":"out there called face ception they're"},{"from":1095.06,"to":1097.46,"location":2,"content":"based in Israel and they claim to be"},{"from":1097.46,"to":1101.93,"location":2,"content":"able to use individual images with"},{"from":1101.93,"to":1103.16,"location":2,"content":"computer vision and machine learning"},{"from":1103.16,"to":1105.26,"location":2,"content":"technology for profiling people and"},{"from":1105.26,"to":1107.39,"location":2,"content":"revealing their personality based only"},{"from":1107.39,"to":1110.96,"location":2,"content":"on their facial image recognizing things"},{"from":1110.96,"to":1113.93,"location":2,"content":"like high IQ white-collar offender had a"},{"from":1113.93,"to":1116.48,"location":2,"content":"file and terrorists and their main"},{"from":1116.48,"to":1118.73,"location":2,"content":"clients are homeland security lots of"},{"from":1118.73,"to":1120.95,"location":2,"content":"other lots of other countries dealing"},{"from":1120.95,"to":1122.42,"location":2,"content":"with sort of public safety issues"},{"from":1122.42,"to":1124.88,"location":2,"content":"they've not published any details about"},{"from":1124.88,"to":1126.83,"location":2,"content":"their methods their sources of training"},{"from":1126.83,"to":1129.32,"location":2,"content":"data or their quantitative results we"},{"from":1129.32,"to":1130.91,"location":2,"content":"know that in light of automation bias"},{"from":1130.91,"to":1132.89,"location":2,"content":"people will tend to think it just works"},{"from":1132.89,"to":1135.71,"location":2,"content":"even when it doesn't work well but there"},{"from":1135.71,"to":1137.75,"location":2,"content":"was a paper that came out within a"},{"from":1137.75,"to":1140.53,"location":2,"content":"similar line in predicting criminal"},{"from":1140.53,"to":1143.09,"location":2,"content":"criminality or purporting to predict"},{"from":1143.09,"to":1145.49,"location":2,"content":"criminality from individual face images"},{"from":1145.49,"to":1149.06,"location":2,"content":"and that one had some results and some"},{"from":1149.06,"to":1150.65,"location":2,"content":"more details about the data that we"},{"from":1150.65,"to":1152.3,"location":2,"content":"could kind of dig into to understand"},{"from":1152.3,"to":1154.49,"location":2,"content":"where are these kinds of claims coming"},{"from":1154.49,"to":1156.71,"location":2,"content":"from so this was an article that was"},{"from":1156.71,"to":1159.52,"location":2,"content":"posted on archive near the end of 2016"},{"from":1159.52,"to":1162.32,"location":2,"content":"and they said they were using less than"},{"from":1162.32,"to":1165.52,"location":2,"content":"2000 closely cropped images of faces"},{"from":1165.52,"to":1168.71,"location":2,"content":"including wanted suspect ID pictures"},{"from":1168.71,"to":1171.14,"location":2,"content":"from specific regions and they claimed"},{"from":1171.14,"to":1172.84,"location":2,"content":"that even based on this very small"},{"from":1172.84,"to":1176.27,"location":2,"content":"training data set that they were able to"},{"from":1176.27,"to":1178.49,"location":2,"content":"predict whether or not someone was"},{"from":1178.49,"to":1181.46,"location":2,"content":"likely to be a criminal greater than 90"},{"from":1181.46,"to":1182.36,"location":2,"content":"percent accuracy"},{"from":1182.36,"to":1185.51,"location":2,"content":"um and they got so lost in this this"},{"from":1185.51,"to":1188.72,"location":2,"content":"idea that it's sort of funny to read to"},{"from":1188.72,"to":1190.4,"location":2,"content":"just take a step back and realize what's"},{"from":1190.4,"to":1192.86,"location":2,"content":"actually happening so for example one of"},{"from":1192.86,"to":1195.47,"location":2,"content":"their really great exciting claims was"},{"from":1195.47,"to":1197.84,"location":2,"content":"that the angle theta from nose tip 2 to"},{"from":1197.84,"to":1200.15,"location":2,"content":"mouth corners is on average nineteen"},{"from":1200.15,"to":1202.49,"location":2,"content":"point six percent smaller for criminals"},{"from":1202.49,"to":1204.89,"location":2,"content":"for non-criminals this is otherwise"},{"from":1204.89,"to":1209.57,"location":2,"content":"known as smiling and my you know exactly"},{"from":1209.57,"to":1211.52,"location":2,"content":"the kind of images people would use when"},{"from":1211.52,"to":1213.2,"location":2,"content":"trying to put out wanted criminal"},{"from":1213.2,"to":1214.79,"location":2,"content":"pictures probably not really happy"},{"from":1214.79,"to":1216.98,"location":2,"content":"pictures but you get so lost in the"},{"from":1216.98,"to":1219.29,"location":2,"content":"confirmation bias you get so lost in the"},{"from":1219.29,"to":1221.39,"location":2,"content":"correlation and the feedback loops that"},{"from":1221.39,"to":1223.37,"location":2,"content":"you end up overlooking these really"},{"from":1223.37,"to":1227.6,"location":2,"content":"obvious kinds of things so that's an"},{"from":1227.6,"to":1229.7,"location":2,"content":"example of selection bias experimenters"},{"from":1229.7,"to":1232.31,"location":2,"content":"bias confirmation bias correlation"},{"from":1232.31,"to":1234.29,"location":2,"content":"fallacy and feedback loops all coming"},{"from":1234.29,"to":1236.75,"location":2,"content":"together to create a deep learning"},{"from":1236.75,"to":1238.52,"location":2,"content":"system that people think is scary and"},{"from":1238.52,"to":1240.97,"location":2,"content":"can do things that it can't actually do"},{"from":1240.97,"to":1243.44,"location":2,"content":"one of the issues with this was that the"},{"from":1243.44,"to":1245.63,"location":2,"content":"media loved it like it's was all over"},{"from":1245.63,"to":1247.28,"location":2,"content":"the news and there's been similar kinds"},{"from":1247.28,"to":1248.57,"location":2,"content":"of things happening again and again"},{"from":1248.57,"to":1251.54,"location":2,"content":"media wants to sell this story and so"},{"from":1251.54,"to":1253.94,"location":2,"content":"it's part of our job as researchers that"},{"from":1253.94,"to":1255.86,"location":2,"content":"people who work on this stuff to be very"},{"from":1255.86,"to":1257.48,"location":2,"content":"clear about what the technology is"},{"from":1257.48,"to":1260.03,"location":2,"content":"actually doing and make a distinction"},{"from":1260.03,"to":1261.68,"location":2,"content":"between what you might think it's doing"},{"from":1261.68,"to":1264.68,"location":2,"content":"and what it's actually doing um so"},{"from":1264.68,"to":1267.01,"location":2,"content":"another issue that has come up recently"},{"from":1267.01,"to":1269.42,"location":2,"content":"is claiming to be able to predict"},{"from":1269.42,"to":1271.67,"location":2,"content":"internal qualities but specifically ones"},{"from":1271.67,"to":1273.41,"location":2,"content":"that are subject to discrimination and"},{"from":1273.41,"to":1276.77,"location":2,"content":"loss of opportunity so in particular"},{"from":1276.77,"to":1278.3,"location":2,"content":"there was this work that came out that"},{"from":1278.3,"to":1279.98,"location":2,"content":"claimed to be able to predict whether or"},{"from":1279.98,"to":1282.44,"location":2,"content":"not someone was homosexual just based on"},{"from":1282.44,"to":1285.23,"location":2,"content":"single face images now it's important to"},{"from":1285.23,"to":1287.42,"location":2,"content":"know that the images that they used in"},{"from":1287.42,"to":1289.49,"location":2,"content":"the study included images that were from"},{"from":1289.49,"to":1291.28,"location":2,"content":"dating websites where people"},{"from":1291.28,"to":1293.33,"location":2,"content":"self-identified as straight or gay and"},{"from":1293.33,"to":1295.1,"location":2,"content":"identified as whether they were looking"},{"from":1295.1,"to":1296.93,"location":2,"content":"for a partner who is straight or gay and"},{"from":1296.93,"to":1299.42,"location":2,"content":"these became the sources of the training"},{"from":1299.42,"to":1302.51,"location":2,"content":"data and still from this Oh before I go"},{"from":1302.51,"to":1304.7,"location":2,"content":"on can you guys just understand just"},{"from":1304.7,"to":1307.3,"location":2,"content":"from that what the issue might have been"},{"from":1307.3,"to":1311.82,"location":2,"content":"imposed"},{"from":1311.82,"to":1313.75,"location":2,"content":"there was actually anything about"},{"from":1313.75,"to":1319.23,"location":2,"content":"rainbows but that's really unfortunate"},{"from":1319.23,"to":1321.97,"location":2,"content":"right yes this has more to do with the"},{"from":1321.97,"to":1323.68,"location":2,"content":"presentation of the self the"},{"from":1323.68,"to":1325.33,"location":2,"content":"presentation of the social self when"},{"from":1325.33,"to":1327.46,"location":2,"content":"you're trying to for example attract a"},{"from":1327.46,"to":1329.74,"location":2,"content":"partner on a website and less to do with"},{"from":1329.74,"to":1332.89,"location":2,"content":"how you look day-to-day and yet they"},{"from":1332.89,"to":1336.37,"location":2,"content":"kind of went to these large conclusions"},{"from":1336.37,"to":1338.62,"location":2,"content":"that aren't supported at all by the data"},{"from":1338.62,"to":1340.72,"location":2,"content":"or by their study but things like"},{"from":1340.72,"to":1343.18,"location":2,"content":"consistent with a prenatal formula of"},{"from":1343.18,"to":1345.61,"location":2,"content":"sexual orientation gay men and women"},{"from":1345.61,"to":1347.56,"location":2,"content":"tended to have gender atypical facial"},{"from":1347.56,"to":1349.96,"location":2,"content":"morphology now none of the authors"},{"from":1349.96,"to":1352.51,"location":2,"content":"actually were prenatal hormone Theory"},{"from":1352.51,"to":1355.57,"location":2,"content":"specialists you know they had doctor in"},{"from":1355.57,"to":1356.74,"location":2,"content":"their name so maybe that's the thing"},{"from":1356.74,"to":1359.11,"location":2,"content":"this was a Stanford professor and like"},{"from":1359.11,"to":1361.12,"location":2,"content":"I've presented this a few times at"},{"from":1361.12,"to":1362.47,"location":2,"content":"Stanford and gotten into some like"},{"from":1362.47,"to":1364.96,"location":2,"content":"pretty harsh fights about this so I'm"},{"from":1364.96,"to":1368.34,"location":2,"content":"ready if anyone wants to take me on but"},{"from":1368.34,"to":1371.53,"location":2,"content":"but me and my some of my colleagues"},{"from":1371.53,"to":1373.45,"location":2,"content":"decided we we play around with this a"},{"from":1373.45,"to":1375.67,"location":2,"content":"bit what we found was that a simple"},{"from":1375.67,"to":1377.98,"location":2,"content":"decision tree so I'm kind of assuming"},{"from":1377.98,"to":1379.86,"location":2,"content":"you guys know what a decision tree is"},{"from":1379.86,"to":1383.14,"location":2,"content":"okay cool so based on wearing makeup or"},{"from":1383.14,"to":1385.33,"location":2,"content":"wearing glasses God is pretty close to"},{"from":1385.33,"to":1387.34,"location":2,"content":"the accuracy reported in the paper"},{"from":1387.34,"to":1389.14,"location":2,"content":"that says nothing about internal"},{"from":1389.14,"to":1390.73,"location":2,"content":"hormones that says nothing about any of"},{"from":1390.73,"to":1392.89,"location":2,"content":"that and says a lot about the physical"},{"from":1392.89,"to":1394.72,"location":2,"content":"presentation the things that are on the"},{"from":1394.72,"to":1397.33,"location":2,"content":"surface it says a lot more about how"},{"from":1397.33,"to":1399.04,"location":2,"content":"people are presenting themselves and"},{"from":1399.04,"to":1402.25,"location":2,"content":"what is happening internally so the key"},{"from":1402.25,"to":1403.66,"location":2,"content":"thing that's recently kind of been"},{"from":1403.66,"to":1405.79,"location":2,"content":"overlooked is that deep learning is"},{"from":1405.79,"to":1407.89,"location":2,"content":"somehow it's sort of considered that"},{"from":1407.89,"to":1409.78,"location":2,"content":"it's somehow magically going beyond"},{"from":1409.78,"to":1411.91,"location":2,"content":"surface level but the point is that it's"},{"from":1411.91,"to":1413.65,"location":2,"content":"working on the surface level and working"},{"from":1413.65,"to":1415.9,"location":2,"content":"well and in the face of confirmation"},{"from":1415.9,"to":1417.94,"location":2,"content":"bias and other kinds of bias factors"},{"from":1417.94,"to":1419.98,"location":2,"content":"it's easy to assume that something else"},{"from":1419.98,"to":1422.44,"location":2,"content":"is happening that's not without critical"},{"from":1422.44,"to":1425.97,"location":2,"content":"examination for example simple baselines"},{"from":1425.97,"to":1428.83,"location":2,"content":"simple sanity checks and these kinds of"},{"from":1428.83,"to":1430.93,"location":2,"content":"things can just be ignored and and not"},{"from":1430.93,"to":1435.12,"location":2,"content":"noticed at all so that's example of"},{"from":1435.12,"to":1437.86,"location":2,"content":"selection bias and experimenters bias"},{"from":1437.86,"to":1441.37,"location":2,"content":"and correlation fallacy"},{"from":1441.37,"to":1443.89,"location":2,"content":"okay so now I'm going to talk to talk"},{"from":1443.89,"to":1446.08,"location":2,"content":"about measuring algorithmic bias so I"},{"from":1446.08,"to":1448.42,"location":2,"content":"just said a lot about different kinds of"},{"from":1448.42,"to":1450.91,"location":2,"content":"biases that come in in the data in the"},{"from":1450.91,"to":1453.25,"location":2,"content":"collection in the interpretation of the"},{"from":1453.25,"to":1454.87,"location":2,"content":"results let's talk about actually"},{"from":1454.87,"to":1457.03,"location":2,"content":"quantitatively measuring different kinds"},{"from":1457.03,"to":1460.69,"location":2,"content":"of biases so one of the key things"},{"from":1460.69,"to":1463.42,"location":2,"content":"that's emerged in a few different works"},{"from":1463.42,"to":1465.82,"location":2,"content":"and really ties nicely to a lot of"},{"from":1465.82,"to":1467.41,"location":2,"content":"fairness work is this idea of"},{"from":1467.41,"to":1470.26,"location":2,"content":"disaggregated evaluation so in"},{"from":1470.26,"to":1472.6,"location":2,"content":"disaggregated evaluation you evaluate"},{"from":1472.6,"to":1474.76,"location":2,"content":"across different subgroups as opposed to"},{"from":1474.76,"to":1477.43,"location":2,"content":"looking at one single score for your"},{"from":1477.43,"to":1481.78,"location":2,"content":"overall testing data set so okay you"},{"from":1481.78,"to":1482.83,"location":2,"content":"guys are probably familiar with the"},{"from":1482.83,"to":1484.57,"location":2,"content":"training testing data split you kind of"},{"from":1484.57,"to":1486.88,"location":2,"content":"train on there on your given training"},{"from":1486.88,"to":1488.89,"location":2,"content":"data you test on your given testing data"},{"from":1488.89,"to":1490.92,"location":2,"content":"and you point you report like precision"},{"from":1490.92,"to":1494.29,"location":2,"content":"recall F score or things like that but"},{"from":1494.29,"to":1496.81,"location":2,"content":"what that masks is how well the system"},{"from":1496.81,"to":1498.31,"location":2,"content":"is actually working across different"},{"from":1498.31,"to":1499.72,"location":2,"content":"kinds of individuals and across"},{"from":1499.72,"to":1502.75,"location":2,"content":"different different subgroups and so one"},{"from":1502.75,"to":1504.79,"location":2,"content":"just straightforward way to handle this"},{"from":1504.79,"to":1507.07,"location":2,"content":"is to actually evaluate with respect to"},{"from":1507.07,"to":1509.14,"location":2,"content":"those different subgroups so creating"},{"from":1509.14,"to":1510.67,"location":2,"content":"for each sort of subgroup prediction"},{"from":1510.67,"to":1513.97,"location":2,"content":"pair so for an example you might look at"},{"from":1513.97,"to":1516.52,"location":2,"content":"women face detection men face detection"},{"from":1516.52,"to":1519.12,"location":2,"content":"and look at how the the error rates are"},{"from":1519.12,"to":1523.9,"location":2,"content":"different or similar um another"},{"from":1523.9,"to":1525.67,"location":2,"content":"important part of this is to look at"},{"from":1525.67,"to":1528.87,"location":2,"content":"things intersectionally combining things"},{"from":1528.87,"to":1531.91,"location":2,"content":"like gender and race at the same time"},{"from":1531.91,"to":1534.79,"location":2,"content":"and seeing how those how the error rates"},{"from":1534.79,"to":1537.31,"location":2,"content":"on those sorts of things change and how"},{"from":1537.31,"to":1538.15,"location":2,"content":"they're different across different"},{"from":1538.15,"to":1541.24,"location":2,"content":"intersections and this is inspired by"},{"from":1541.24,"to":1542.25,"location":2,"content":"kimberlé crenshaw"},{"from":1542.25,"to":1544.75,"location":2,"content":"because she she pioneered intersectional"},{"from":1544.75,"to":1548.89,"location":2,"content":"research in critical race theory and she"},{"from":1548.89,"to":1550.45,"location":2,"content":"discussed the story of emma de Graaff"},{"from":1550.45,"to":1554.64,"location":2,"content":"infeed who was a woman at General Motors"},{"from":1554.64,"to":1556.84,"location":2,"content":"and she claimed that the company is"},{"from":1556.84,"to":1559,"location":2,"content":"hiring practices discriminated against"},{"from":1559,"to":1561.7,"location":2,"content":"black women but in their Court opinion"},{"from":1561.7,"to":1563.74,"location":2,"content":"the judges ruled that General Motors"},{"from":1563.74,"to":1566.83,"location":2,"content":"hired many women for secretarial"},{"from":1566.83,"to":1568.96,"location":2,"content":"positions and many black people her"},{"from":1568.96,"to":1571.54,"location":2,"content":"factory roles and thus they could not"},{"from":1571.54,"to":1574.12,"location":2,"content":"have discriminated against black women"},{"from":1574.12,"to":1576.04,"location":2,"content":"what they failed to do was look at the"},{"from":1576.04,"to":1577.75,"location":2,"content":"intersection of the two and understand"},{"from":1577.75,"to":1579.13,"location":2,"content":"that the experience there might be"},{"from":1579.13,"to":1581.29,"location":2,"content":"fundamentally different than any of the"},{"from":1581.29,"to":1584.08,"location":2,"content":"experiences of either of these sort of"},{"from":1584.08,"to":1587.02,"location":2,"content":"subgroups in isolation and the same"},{"from":1587.02,"to":1589.36,"location":2,"content":"becomes true when you start looking at"},{"from":1589.36,"to":1591.34,"location":2,"content":"errors that are regularly made in deep"},{"from":1591.34,"to":1593.65,"location":2,"content":"learning systems so we've been able to"},{"from":1593.65,"to":1595.33,"location":2,"content":"uncover a lot of different kinds of"},{"from":1595.33,"to":1597.22,"location":2,"content":"unintended errors by looking not only at"},{"from":1597.22,"to":1600.07,"location":2,"content":"the disaggregated evaluation but also at"},{"from":1600.07,"to":1603.78,"location":2,"content":"intersectional disaggregated evaluation"},{"from":1603.78,"to":1605.86,"location":2,"content":"so I'm going to walk through a bit how"},{"from":1605.86,"to":1607.96,"location":2,"content":"this works this is probably going to be"},{"from":1607.96,"to":1610.18,"location":2,"content":"review for most of you but I think it's"},{"from":1610.18,"to":1611.56,"location":2,"content":"really important to understand this"},{"from":1611.56,"to":1613.72,"location":2,"content":"because it also ties to how we measure"},{"from":1613.72,"to":1615.66,"location":2,"content":"fairness and when we say like"},{"from":1615.66,"to":1617.74,"location":2,"content":"algorithmic fairness what we're talking"},{"from":1617.74,"to":1621.79,"location":2,"content":"about so um the confusion matrix is a"},{"from":1621.79,"to":1623.59,"location":2,"content":"way you guys okay are you guys familiar"},{"from":1623.59,"to":1625.78,"location":2,"content":"with the Confucian matrix I just want to"},{"from":1625.78,"to":1628.21,"location":2,"content":"know where okay awesome cool so familiar"},{"from":1628.21,"to":1629.29,"location":2,"content":"to take from you is a matrix right so"},{"from":1629.29,"to":1630.46,"location":2,"content":"you have model predictions and"},{"from":1630.46,"to":1632.86,"location":2,"content":"references and you can kind of look at"},{"from":1632.86,"to":1635.26,"location":2,"content":"these as negative and positive binary"},{"from":1635.26,"to":1637.93,"location":2,"content":"classification kind of approach here"},{"from":1637.93,"to":1641.23,"location":2,"content":"where if the ground truth says something"},{"from":1641.23,"to":1643.12,"location":2,"content":"is true and the model predicts it's true"},{"from":1643.12,"to":1645.13,"location":2,"content":"it's a true positive if the ground truth"},{"from":1645.13,"to":1649.06,"location":2,"content":"says it's it's it's false and the model"},{"from":1649.06,"to":1650.49,"location":2,"content":"predicts it's false it's true negative"},{"from":1650.49,"to":1653.26,"location":2,"content":"and the air is the kind of different"},{"from":1653.26,"to":1654.91,"location":2,"content":"issues that arise are false negatives"},{"from":1654.91,"to":1657.52,"location":2,"content":"and false positives so in false"},{"from":1657.52,"to":1660.91,"location":2,"content":"positives the the ground truth says"},{"from":1660.91,"to":1662.74,"location":2,"content":"something is negative but the model"},{"from":1662.74,"to":1665.53,"location":2,"content":"predicts that it's positive and then in"},{"from":1665.53,"to":1667.65,"location":2,"content":"false negatives vice versa"},{"from":1667.65,"to":1671.35,"location":2,"content":"from these you know basic kind of this"},{"from":1671.35,"to":1673.81,"location":2,"content":"basic breakdown of errors you can get a"},{"from":1673.81,"to":1677.2,"location":2,"content":"few different metrics these metrics"},{"from":1677.2,"to":1679.45,"location":2,"content":"actually trivially map to a lot of"},{"from":1679.45,"to":1682.18,"location":2,"content":"different fairness criteria so for"},{"from":1682.18,"to":1684.1,"location":2,"content":"example if we're looking at something"},{"from":1684.1,"to":1686.89,"location":2,"content":"like female versus male patient results"},{"from":1686.89,"to":1689.08,"location":2,"content":"and figuring out things like precision"},{"from":1689.08,"to":1691.06,"location":2,"content":"and recall which is relatively common in"},{"from":1691.06,"to":1695.5,"location":2,"content":"NLP if you have equal recall across your"},{"from":1695.5,"to":1696.16,"location":2,"content":"subgroups"},{"from":1696.16,"to":1698.71,"location":2,"content":"that's the same as the fairness criteria"},{"from":1698.71,"to":1702.79,"location":2,"content":"of equality of opportunity I could work"},{"from":1702.79,"to":1704.05,"location":2,"content":"through the math but I mean this is"},{"from":1704.05,"to":1705.88,"location":2,"content":"basically just just the main point that"},{"from":1705.88,"to":1707.57,"location":2,"content":"that"},{"from":1707.57,"to":1709.85,"location":2,"content":"it says that given that something is"},{"from":1709.85,"to":1713.21,"location":2,"content":"true in the ground truth the model"},{"from":1713.21,"to":1716.09,"location":2,"content":"should predict that it's true at equal"},{"from":1716.09,"to":1717.89,"location":2,"content":"rates across different subgroups so this"},{"from":1717.89,"to":1719.63,"location":2,"content":"ends up being equivalent to having the"},{"from":1719.63,"to":1721.84,"location":2,"content":"same recall across different subgroups"},{"from":1721.84,"to":1724.85,"location":2,"content":"similarly having the same precision"},{"from":1724.85,"to":1727.37,"location":2,"content":"across different subgroups is equivalent"},{"from":1727.37,"to":1729.17,"location":2,"content":"to a fairness criterion called"},{"from":1729.17,"to":1732.65,"location":2,"content":"predictive parity and so as fairness has"},{"from":1732.65,"to":1735.86,"location":2,"content":"been defined again and again it was"},{"from":1735.86,"to":1737.48,"location":2,"content":"originally some of these definitions"},{"from":1737.48,"to":1741.17,"location":2,"content":"came in 1966 following the Civil Rights"},{"from":1741.17,"to":1745.16,"location":2,"content":"Act of 1964 they were reinvented a few"},{"from":1745.16,"to":1748.51,"location":2,"content":"times and most recently reinvented in"},{"from":1748.51,"to":1752.51,"location":2,"content":"2016 but they all sort of boiled down to"},{"from":1752.51,"to":1754.79,"location":2,"content":"this disaggregated comparison across"},{"from":1754.79,"to":1757.31,"location":2,"content":"subgroups and the math the metrics end"},{"from":1757.31,"to":1759.29,"location":2,"content":"up being roughly equivalent to what we"},{"from":1759.29,"to":1760.48,"location":2,"content":"get from the confusion matrix"},{"from":1760.48,"to":1766.03,"location":2,"content":"specifically in classification systems"},{"from":1766.03,"to":1768.95,"location":2,"content":"so which kind of fairness metric do you"},{"from":1768.95,"to":1771.56,"location":2,"content":"use what are the different criteria you"},{"from":1771.56,"to":1773.69,"location":2,"content":"want to use to look at the differences"},{"from":1773.69,"to":1775.61,"location":2,"content":"across different subgroups that really"},{"from":1775.61,"to":1777.86,"location":2,"content":"it comes down to the trade-offs between"},{"from":1777.86,"to":1780.08,"location":2,"content":"false positives and false negatives so"},{"from":1780.08,"to":1781.49,"location":2,"content":"this is the same problem that you're"},{"from":1781.49,"to":1782.63,"location":2,"content":"dealing with when you're just figuring"},{"from":1782.63,"to":1785.24,"location":2,"content":"out how to evaluate generally there's no"},{"from":1785.24,"to":1787.01,"location":2,"content":"one fairness criterion that is the"},{"from":1787.01,"to":1789.43,"location":2,"content":"fairness criterion to rule them all"},{"from":1789.43,"to":1791.39,"location":2,"content":"deciding which one is better than the"},{"from":1791.39,"to":1793.19,"location":2,"content":"other is the same as kind of trying to"},{"from":1793.19,"to":1794.75,"location":2,"content":"decide which is better precision or"},{"from":1794.75,"to":1796.22,"location":2,"content":"recall right it depends on what the"},{"from":1796.22,"to":1797.66,"location":2,"content":"problem is and what you're interested in"},{"from":1797.66,"to":1800.93,"location":2,"content":"measuring so a case where false"},{"from":1800.93,"to":1803.3,"location":2,"content":"positives might be better than false"},{"from":1803.3,"to":1805.55,"location":2,"content":"negatives and so you want to prioritize"},{"from":1805.55,"to":1807.7,"location":2,"content":"something like a false positive right"},{"from":1807.7,"to":1810.92,"location":2,"content":"across subgroups is privacy in images so"},{"from":1810.92,"to":1812.6,"location":2,"content":"here are false positive is something"},{"from":1812.6,"to":1814.76,"location":2,"content":"that doesn't need to be blurred gets"},{"from":1814.76,"to":1817.37,"location":2,"content":"blurred that's just kind of a bummer but"},{"from":1817.37,"to":1818.66,"location":2,"content":"a false negative would be something that"},{"from":1818.66,"to":1821,"location":2,"content":"needs to be blurred as not learned and"},{"from":1821,"to":1822.8,"location":2,"content":"that can be identity theft it's a much"},{"from":1822.8,"to":1825.29,"location":2,"content":"more serious issue and so it's important"},{"from":1825.29,"to":1827.21,"location":2,"content":"to prioritize the evaluation metrics"},{"from":1827.21,"to":1830.78,"location":2,"content":"that stress the false negative rates an"},{"from":1830.78,"to":1832.91,"location":2,"content":"example where false negatives might be"},{"from":1832.91,"to":1834.65,"location":2,"content":"better than false positives as in spam"},{"from":1834.65,"to":1837.26,"location":2,"content":"filtering so a false negative could be"},{"from":1837.26,"to":1838.87,"location":2,"content":"an email that's spam"},{"from":1838.87,"to":1840.84,"location":2,"content":"not caught so you see it in your inbox"},{"from":1840.84,"to":1843.49,"location":2,"content":"that's usually just annoying it's not a"},{"from":1843.49,"to":1845.86,"location":2,"content":"big deal and but if false positive here"},{"from":1845.86,"to":1847.87,"location":2,"content":"would be email flagged as spam and then"},{"from":1847.87,"to":1850.45,"location":2,"content":"removed from your inbox which you know"},{"from":1850.45,"to":1853.18,"location":2,"content":"if it's from a friend or a loved one it"},{"from":1853.18,"to":1855.01,"location":2,"content":"can be it can be a loss may be a job"},{"from":1855.01,"to":1857.16,"location":2,"content":"offer or something like that"},{"from":1857.16,"to":1861.13,"location":2,"content":"right so I just kind of covered how a I"},{"from":1861.13,"to":1863.32,"location":2,"content":"can unintentionally timejust outcomes"},{"from":1863.32,"to":1865.48,"location":2,"content":"and some of the things to do or some of"},{"from":1865.48,"to":1866.89,"location":2,"content":"the things to be aware of here"},{"from":1866.89,"to":1869.29,"location":2,"content":"are the lack of insight into sources of"},{"from":1869.29,"to":1871.08,"location":2,"content":"bias in the data in the model"},{"from":1871.08,"to":1873.85,"location":2,"content":"lack of insight into the feedback loops"},{"from":1873.85,"to":1876.25,"location":2,"content":"from the original data that's collected"},{"from":1876.25,"to":1879.55,"location":2,"content":"as an example of what humans do to the"},{"from":1879.55,"to":1882.52,"location":2,"content":"data that's then repurposed reused acted"},{"from":1882.52,"to":1885.58,"location":2,"content":"on and then further fed in a lack of"},{"from":1885.58,"to":1888.43,"location":2,"content":"careful disaggregated evaluation looking"},{"from":1888.43,"to":1890.2,"location":2,"content":"at the disparities the differences"},{"from":1890.2,"to":1892.33,"location":2,"content":"between different subgroups in order to"},{"from":1892.33,"to":1894.28,"location":2,"content":"understand this bias this difference"},{"from":1894.28,"to":1896.65,"location":2,"content":"across the subgroups and then human"},{"from":1896.65,"to":1898.9,"location":2,"content":"biases in interpreting and accepting and"},{"from":1898.9,"to":1900.94,"location":2,"content":"talking about the results which then"},{"from":1900.94,"to":1903.34,"location":2,"content":"kind of further the media cycles and the"},{"from":1903.34,"to":1908.11,"location":2,"content":"hype around AI right now but it's up to"},{"from":1908.11,"to":1912.34,"location":2,"content":"us to influence how AI evolves so I like"},{"from":1912.34,"to":1914.95,"location":2,"content":"to think of this in terms of short term"},{"from":1914.95,"to":1917.73,"location":2,"content":"middle term and long term objectives so"},{"from":1917.73,"to":1921.76,"location":2,"content":"short term today we might be working on"},{"from":1921.76,"to":1923.68,"location":2,"content":"some specific modal where we're trying"},{"from":1923.68,"to":1925.51,"location":2,"content":"to find some local optimum we have a"},{"from":1925.51,"to":1927.28,"location":2,"content":"task we have data something like that"},{"from":1927.28,"to":1929.97,"location":2,"content":"and that sort of short term objectives"},{"from":1929.97,"to":1932.29,"location":2,"content":"we might have a slightly longer term"},{"from":1932.29,"to":1933.94,"location":2,"content":"objective of getting a paper published"},{"from":1933.94,"to":1936.07,"location":2,"content":"or if you're an industry like adding a"},{"from":1936.07,"to":1939.22,"location":2,"content":"product launched whatever it might be"},{"from":1939.22,"to":1940.66,"location":2,"content":"from there we might see our next"},{"from":1940.66,"to":1943.75,"location":2,"content":"endpoint as getting an award or you know"},{"from":1943.75,"to":1945.37,"location":2,"content":"maybe become sort of famous for"},{"from":1945.37,"to":1946.6,"location":2,"content":"something for a few minutes something"},{"from":1946.6,"to":1949.42,"location":2,"content":"like that and that's cool and but"},{"from":1949.42,"to":1951.28,"location":2,"content":"there's a longer-term objective that we"},{"from":1951.28,"to":1953.32,"location":2,"content":"can work towards as well at the same"},{"from":1953.32,"to":1955.24,"location":2,"content":"time and that's something like a"},{"from":1955.24,"to":1957.19,"location":2,"content":"positive outcome for humans in their"},{"from":1957.19,"to":1959.53,"location":2,"content":"environment so instead of just kind of"},{"from":1959.53,"to":1962.26,"location":2,"content":"focusing on these local decisions these"},{"from":1962.26,"to":1964.36,"location":2,"content":"local optimum and these sort of local"},{"from":1964.36,"to":1967.42,"location":2,"content":"paper by paper based approaches to"},{"from":1967.42,"to":1969.52,"location":2,"content":"solving problems you can also kind of"},{"from":1969.52,"to":1970.78,"location":2,"content":"think about what's the long-term"},{"from":1970.78,"to":1972.23,"location":2,"content":"objective where does this"},{"from":1972.23,"to":1974.81,"location":2,"content":"me as they trace out an evolutionary"},{"from":1974.81,"to":1977.03,"location":2,"content":"path for artificial intelligence down"},{"from":1977.03,"to":1981.94,"location":2,"content":"the line in 10 years 15 years 20 years"},{"from":1981.94,"to":1985.16,"location":2,"content":"and one of the ways you can address this"},{"from":1985.16,"to":1987.14,"location":2,"content":"is by thinking you know how can the work"},{"from":1987.14,"to":1989.57,"location":2,"content":"I'm interested in now be best focused to"},{"from":1989.57,"to":1991.46,"location":2,"content":"help others and that involves talking to"},{"from":1991.46,"to":1993.41,"location":2,"content":"experts and kind of going outside your"},{"from":1993.41,"to":1995.9,"location":2,"content":"bubble speaking across interdisciplinary"},{"from":1995.9,"to":1997.67,"location":2,"content":"fields like cognitive science which I've"},{"from":1997.67,"to":2001.45,"location":2,"content":"just talked a bit about so let's talk"},{"from":2001.45,"to":2004.72,"location":2,"content":"about some things we can do so first off"},{"from":2004.72,"to":2010.6,"location":2,"content":"is data so a lot of the issues of bias"},{"from":2010.6,"to":2014.05,"location":2,"content":"and fairness in machine learning models"},{"from":2014.05,"to":2015.79,"location":2,"content":"really come down to the data"},{"from":2015.79,"to":2018.25,"location":2,"content":"unfortunately in machine learning and"},{"from":2018.25,"to":2021.31,"location":2,"content":"deep learning working on data is really"},{"from":2021.31,"to":2024.49,"location":2,"content":"not seen as sexy there's a few data sets"},{"from":2024.49,"to":2027.64,"location":2,"content":"that people use they're out there that's"},{"from":2027.64,"to":2029.47,"location":2,"content":"what people use and there's not a lot of"},{"from":2029.47,"to":2031.78,"location":2,"content":"analysis done on it on how well these"},{"from":2031.78,"to":2034.21,"location":2,"content":"datasets capture different truths about"},{"from":2034.21,"to":2038.49,"location":2,"content":"the world how problematic they might be"},{"from":2038.49,"to":2041.23,"location":2,"content":"but it's a pretty wide area that needs a"},{"from":2041.23,"to":2043.42,"location":2,"content":"lot of future like lead needs a lot of"},{"from":2043.42,"to":2046.18,"location":2,"content":"future additional work so we're gonna"},{"from":2046.18,"to":2047.74,"location":2,"content":"understand the data skews and the"},{"from":2047.74,"to":2049.75,"location":2,"content":"correlations if you understand your data"},{"from":2049.75,"to":2052.63,"location":2,"content":"skews and the correlations that might be"},{"from":2052.63,"to":2054.49,"location":2,"content":"problematic in your data then you can"},{"from":2054.49,"to":2056.32,"location":2,"content":"start working on either models that"},{"from":2056.32,"to":2058.78,"location":2,"content":"address those or data augmentation"},{"from":2058.78,"to":2061.06,"location":2,"content":"approaches in order to sort of make the"},{"from":2061.06,"to":2062.71,"location":2,"content":"data set a little bit better or a little"},{"from":2062.71,"to":2064.69,"location":2,"content":"bit more representative of how you want"},{"from":2064.69,"to":2067.75,"location":2,"content":"the world to be it's also important to"},{"from":2067.75,"to":2070.24,"location":2,"content":"abandon the single training set testing"},{"from":2070.24,"to":2072.28,"location":2,"content":"set from similar distribution approach"},{"from":2072.28,"to":2076.84,"location":2,"content":"to advancing deep learning so when we do"},{"from":2076.84,"to":2078.31,"location":2,"content":"projects in deep learning you know we"},{"from":2078.31,"to":2079.78,"location":2,"content":"tend to have the training set and the"},{"from":2079.78,"to":2081.43,"location":2,"content":"testing set and then that's what we sort"},{"from":2081.43,"to":2083.65,"location":2,"content":"of benchmark on and prioritize but the"},{"from":2083.65,"to":2085.3,"location":2,"content":"point is as you move around different"},{"from":2085.3,"to":2086.98,"location":2,"content":"testing sets you're going to get vastly"},{"from":2086.98,"to":2090.37,"location":2,"content":"different results and so by keeping in"},{"from":2090.37,"to":2092.26,"location":2,"content":"this just sort of one training testing"},{"from":2092.26,"to":2095.44,"location":2,"content":"date training testing data set paradigm"},{"from":2095.44,"to":2097.69,"location":2,"content":"you're really likely to not notice"},{"from":2097.69,"to":2100.15,"location":2,"content":"issues that might otherwise be there and"},{"from":2100.15,"to":2102.25,"location":2,"content":"one way to really focus in on them is"},{"from":2102.25,"to":2106.57,"location":2,"content":"having a hard set of test cases"},{"from":2106.57,"to":2108.16,"location":2,"content":"you really want to make sure the model"},{"from":2108.16,"to":2109.87,"location":2,"content":"does well on so these are things that"},{"from":2109.87,"to":2112.57,"location":2,"content":"are particularly problematic things that"},{"from":2112.57,"to":2114.73,"location":2,"content":"would be really harmful to individuals"},{"from":2114.73,"to":2117.16,"location":2,"content":"if they were to experience the output"},{"from":2117.16,"to":2119.71,"location":2,"content":"and you kind of collect those in a small"},{"from":2119.71,"to":2121.66,"location":2,"content":"test set and then it's really easy to"},{"from":2121.66,"to":2124.36,"location":2,"content":"evaluate on that test set as you"},{"from":2124.36,"to":2126.1,"location":2,"content":"benchmark improvements on your model as"},{"from":2126.1,"to":2127.72,"location":2,"content":"you add different kinds of things to"},{"from":2127.72,"to":2130.6,"location":2,"content":"your model in order to see not just how"},{"from":2130.6,"to":2132.67,"location":2,"content":"your model is doing overall in terms of"},{"from":2132.67,"to":2134.56,"location":2,"content":"your testing data set but how well"},{"from":2134.56,"to":2136.48,"location":2,"content":"you're doing in terms of these examples"},{"from":2136.48,"to":2138.73,"location":2,"content":"you really want it to do well on that"},{"from":2138.73,"to":2140.65,"location":2,"content":"you know that it's going to be a problem"},{"from":2140.65,"to":2142.72,"location":2,"content":"if it doesn't do well on and any sort of"},{"from":2142.72,"to":2144.46,"location":2,"content":"degradation in that you might want to"},{"from":2144.46,"to":2147.97,"location":2,"content":"prioritize to fix above degradation"},{"from":2147.97,"to":2151.57,"location":2,"content":"degradation and overall accuracy and"},{"from":2151.57,"to":2153.25,"location":2,"content":"it's also important to talk to experts"},{"from":2153.25,"to":2154.96,"location":2,"content":"about the additional signals that you"},{"from":2154.96,"to":2159.52,"location":2,"content":"can incorporate so we've put out a tool"},{"from":2159.52,"to":2161.59,"location":2,"content":"to help with this understanding data"},{"from":2161.59,"to":2163.99,"location":2,"content":"SKUs called facets it's just available"},{"from":2163.99,"to":2166.81,"location":2,"content":"there and it's a really handy kind of"},{"from":2166.81,"to":2170.86,"location":2,"content":"visualizer for slicing understanding you"},{"from":2170.86,"to":2171.97,"location":2,"content":"know what some of the differences are"},{"from":2171.97,"to":2173.38,"location":2,"content":"between different subgroups and"},{"from":2173.38,"to":2174.73,"location":2,"content":"different representations and you can"},{"from":2174.73,"to":2176.83,"location":2,"content":"sort of dig in and explore a bit more so"},{"from":2176.83,"to":2178.93,"location":2,"content":"this is just to sort of help people come"},{"from":2178.93,"to":2180.16,"location":2,"content":"to terms with the data that they're"},{"from":2180.16,"to":2182.08,"location":2,"content":"actually using and and where there might"},{"from":2182.08,"to":2185.53,"location":2,"content":"be unwanted associations or or missing"},{"from":2185.53,"to":2190.54,"location":2,"content":"missing kind of features another"},{"from":2190.54,"to":2192.31,"location":2,"content":"approach that's been put forward"},{"from":2192.31,"to":2195.34,"location":2,"content":"recently specifically on the data side"},{"from":2195.34,"to":2197.8,"location":2,"content":"is this data data sheets for data sets"},{"from":2197.8,"to":2200.65,"location":2,"content":"approach so this is this idea that when"},{"from":2200.65,"to":2202.84,"location":2,"content":"you release a data set it's not enough"},{"from":2202.84,"to":2204.7,"location":2,"content":"to just release the data set with like"},{"from":2204.7,"to":2206.83,"location":2,"content":"some pretty graphs and like talking"},{"from":2206.83,"to":2208.6,"location":2,"content":"about basic distributional information"},{"from":2208.6,"to":2210.22,"location":2,"content":"you need to talk about who the"},{"from":2210.22,"to":2212.65,"location":2,"content":"annotators were where they were what the"},{"from":2212.65,"to":2214.63,"location":2,"content":"inner entertainer agreement was what"},{"from":2214.63,"to":2216.78,"location":2,"content":"their background information was"},{"from":2216.78,"to":2219.28,"location":2,"content":"motivation for the data set all these"},{"from":2219.28,"to":2220.9,"location":2,"content":"other kinds of details so now you"},{"from":2220.9,"to":2222.7,"location":2,"content":"actually know that this isn't just a"},{"from":2222.7,"to":2224.89,"location":2,"content":"data set this is a data set that has"},{"from":2224.89,"to":2227.35,"location":2,"content":"these specific biases there's no such"},{"from":2227.35,"to":2229.36,"location":2,"content":"thing as a data set that isn't biased in"},{"from":2229.36,"to":2231.73,"location":2,"content":"some way a data set by virtue of the"},{"from":2231.73,"to":2233.44,"location":2,"content":"fact that it's collected from the world"},{"from":2233.44,"to":2237.22,"location":2,"content":"as a subset is a is a biased set of the"},{"from":2237.22,"to":2239.23,"location":2,"content":"world in some way the point is to make"},{"from":2239.23,"to":2239.83,"location":2,"content":"it clear"},{"from":2239.83,"to":2241.81,"location":2,"content":"what it is how it is biased what are the"},{"from":2241.81,"to":2244,"location":2,"content":"what are the various biases that's that"},{"from":2244,"to":2245.41,"location":2,"content":"important to know about in the data set"},{"from":2245.41,"to":2247.06,"location":2,"content":"so that's one of these ideas between"},{"from":2247.06,"to":2249.28,"location":2,"content":"behind data sheets for data sets"},{"from":2249.28,"to":2253.12,"location":2,"content":"releasing these data sets publicly all"},{"from":2253.12,"to":2254.41,"location":2,"content":"right now let's switch a little bit to"},{"from":2254.41,"to":2257.62,"location":2,"content":"machine learning so there are a couple"},{"from":2257.62,"to":2259.57,"location":2,"content":"techniques that I like to use I'll talk"},{"from":2259.57,"to":2262.99,"location":2,"content":"about two one is bias mitigation which"},{"from":2262.99,"to":2264.79,"location":2,"content":"is removing the signal for a problematic"},{"from":2264.79,"to":2269.14,"location":2,"content":"output so removing stereotyping sexism"},{"from":2269.14,"to":2271.09,"location":2,"content":"racism trying to remove these kind of"},{"from":2271.09,"to":2273.19,"location":2,"content":"effects from the model this is also"},{"from":2273.19,"to":2275.85,"location":2,"content":"sometimes called D biasing or unbiased"},{"from":2275.85,"to":2278.62,"location":2,"content":"thats a little bit of a misnomer because"},{"from":2278.62,"to":2280.48,"location":2,"content":"you're you're generally just kind of"},{"from":2280.48,"to":2283.03,"location":2,"content":"moving around bias based on a specific"},{"from":2283.03,"to":2285.58,"location":2,"content":"set of words for example so to say it's"},{"from":2285.58,"to":2287.95,"location":2,"content":"unbiased is it's not true"},{"from":2287.95,"to":2289.78,"location":2,"content":"but you are kind of mitigating bias with"},{"from":2289.78,"to":2291.4,"location":2,"content":"respect to some certain kinds of"},{"from":2291.4,"to":2295.18,"location":2,"content":"information that you provide it with and"},{"from":2295.18,"to":2296.92,"location":2,"content":"there's inclusion which is then adding"},{"from":2296.92,"to":2299.32,"location":2,"content":"signal for desired variables so that's"},{"from":2299.32,"to":2301.09,"location":2,"content":"kind of the opposite side of bias"},{"from":2301.09,"to":2303.28,"location":2,"content":"mitigation so increasing model"},{"from":2303.28,"to":2305.08,"location":2,"content":"performance with attention to subgroups"},{"from":2305.08,"to":2306.7,"location":2,"content":"or data slices with the worst"},{"from":2306.7,"to":2312.4,"location":2,"content":"performance so in order to address"},{"from":2312.4,"to":2315.31,"location":2,"content":"inclusion kind of adding signal for"},{"from":2315.31,"to":2317.53,"location":2,"content":"underrepresented subgroups one technique"},{"from":2317.53,"to":2318.85,"location":2,"content":"that's worked relatively well is"},{"from":2318.85,"to":2321.22,"location":2,"content":"multitask learning so I've heard that"},{"from":2321.22,"to":2323.17,"location":2,"content":"you guys have studied multitask learning"},{"from":2323.17,"to":2324.85,"location":2,"content":"which is great so I'll tell you a bit"},{"from":2324.85,"to":2327.25,"location":2,"content":"about a case study here and so this is"},{"from":2327.25,"to":2330.07,"location":2,"content":"work I did in collaboration with a UPenn"},{"from":2330.07,"to":2332.74,"location":2,"content":"world well-being project working"},{"from":2332.74,"to":2334.72,"location":2,"content":"directly with clinicians and the goal"},{"from":2334.72,"to":2336.43,"location":2,"content":"was to create a system that could alert"},{"from":2336.43,"to":2337.96,"location":2,"content":"clinicians if there was a suicide"},{"from":2337.96,"to":2340.51,"location":2,"content":"attempt that was imminent and they"},{"from":2340.51,"to":2342.31,"location":2,"content":"wanted to understand the feasibility of"},{"from":2342.31,"to":2344.08,"location":2,"content":"these kinds of diagnosis when there were"},{"from":2344.08,"to":2346.9,"location":2,"content":"very few training training instances"},{"from":2346.9,"to":2348.82,"location":2,"content":"available so that's similar to kind of"},{"from":2348.82,"to":2355.72,"location":2,"content":"the minority problem in datasets and in"},{"from":2355.72,"to":2359.26,"location":2,"content":"this work we had two kinds of data one"},{"from":2359.26,"to":2360.88,"location":2,"content":"was the internal data which was the"},{"from":2360.88,"to":2364.12,"location":2,"content":"electronic health records with that was"},{"from":2364.12,"to":2366.07,"location":2,"content":"either provided by the patient or from"},{"from":2366.07,"to":2367.57,"location":2,"content":"the family"},{"from":2367.57,"to":2369.72,"location":2,"content":"it included mental health diagnosis"},{"from":2369.72,"to":2372.46,"location":2,"content":"suicide attempts or completions"},{"from":2372.46,"to":2375.43,"location":2,"content":"if that were the case along with the the"},{"from":2375.43,"to":2377.52,"location":2,"content":"users the person's social media data"},{"from":2377.52,"to":2379.81,"location":2,"content":"that was the internal data that we did"},{"from":2379.81,"to":2381.58,"location":2,"content":"not publish on but that we were able to"},{"from":2381.58,"to":2383.29,"location":2,"content":"work with clinicians on in order to"},{"from":2383.29,"to":2384.79,"location":2,"content":"understand if our methods were actually"},{"from":2384.79,"to":2388.18,"location":2,"content":"working the external data the proxy data"},{"from":2388.18,"to":2389.71,"location":2,"content":"the stuff that we could kind of publish"},{"from":2389.71,"to":2391.86,"location":2,"content":"on and talk about was based on Twitter"},{"from":2391.86,"to":2395.29,"location":2,"content":"and this was using regular expressions"},{"from":2395.29,"to":2399.13,"location":2,"content":"in order to extract phrases and Twitter"},{"from":2399.13,"to":2400.87,"location":2,"content":"feeds that had something that was kind"},{"from":2400.87,"to":2403.39,"location":2,"content":"of like diagnosis so something like I've"},{"from":2403.39,"to":2405.58,"location":2,"content":"been diagnosed with X or I've tried to"},{"from":2405.58,"to":2408.16,"location":2,"content":"commit suicide and that became kind of"},{"from":2408.16,"to":2410.38,"location":2,"content":"the the proxy data set and the"},{"from":2410.38,"to":2412.51,"location":2,"content":"corresponding social media feeds for for"},{"from":2412.51,"to":2414.58,"location":2,"content":"those individuals for the actual"},{"from":2414.58,"to":2420.46,"location":2,"content":"diagnosis and the state of the art in"},{"from":2420.46,"to":2423.58,"location":2,"content":"clinical medicine kind of until this"},{"from":2423.58,"to":2426.04,"location":2,"content":"work there's been more recently but is"},{"from":2426.04,"to":2428.35,"location":2,"content":"sort of this single task logistic"},{"from":2428.35,"to":2430.48,"location":2,"content":"regression logistic regression setup"},{"from":2430.48,"to":2432.19,"location":2,"content":"where you have some input features and"},{"from":2432.19,"to":2433.24,"location":2,"content":"then you're making some output"},{"from":2433.24,"to":2436.99,"location":2,"content":"predictions like true or false and you"},{"from":2436.99,"to":2438.91,"location":2,"content":"can add some layers and start making it"},{"from":2438.91,"to":2442.78,"location":2,"content":"deep learning which is much fancier you"},{"from":2442.78,"to":2445.45,"location":2,"content":"can have a bunch of tasks in order to do"},{"from":2445.45,"to":2447.67,"location":2,"content":"a bunch of logistic regression tasks for"},{"from":2447.67,"to":2450.43,"location":2,"content":"a clinical environment or you can use"},{"from":2450.43,"to":2452.89,"location":2,"content":"multi task learning which is taking a"},{"from":2452.89,"to":2454.54,"location":2,"content":"basic deep learning model and adding a"},{"from":2454.54,"to":2456.88,"location":2,"content":"bunch of heads to it predicted jointly"},{"from":2456.88,"to":2460.15,"location":2,"content":"at the same time and here we had a bunch"},{"from":2460.15,"to":2463.39,"location":2,"content":"of diagnosis data so we predicted things"},{"from":2463.39,"to":2465.85,"location":2,"content":"like depression anxiety post-traumatic"},{"from":2465.85,"to":2469.93,"location":2,"content":"stress disorder we also added in gender"},{"from":2469.93,"to":2471.94,"location":2,"content":"because this is something that the"},{"from":2471.94,"to":2474.37,"location":2,"content":"clinicians told us actually had some"},{"from":2474.37,"to":2475.57,"location":2,"content":"correlation with some of these"},{"from":2475.57,"to":2477.04,"location":2,"content":"conditions and that they actually used"},{"from":2477.04,"to":2479.11,"location":2,"content":"it in making decisions themselves for"},{"from":2479.11,"to":2481.26,"location":2,"content":"whether or not someone was likely to"},{"from":2481.26,"to":2485.44,"location":2,"content":"attempt suicide or not and this also"},{"from":2485.44,"to":2487.45,"location":2,"content":"used this idea of comorbidity so"},{"from":2487.45,"to":2489.61,"location":2,"content":"multitask learning is actually kind of"},{"from":2489.61,"to":2492.31,"location":2,"content":"perfect for comorbidity in clinical"},{"from":2492.31,"to":2495.22,"location":2,"content":"domains so comorbidity is and when you"},{"from":2495.22,"to":2496.63,"location":2,"content":"have one condition you're a lot more"},{"from":2496.63,"to":2499.51,"location":2,"content":"likely to have another so people who"},{"from":2499.51,"to":2501.1,"location":2,"content":"have post-traumatic stress disorder are"},{"from":2501.1,"to":2502.78,"location":2,"content":"much more likely to have depression and"},{"from":2502.78,"to":2505.42,"location":2,"content":"anxiety and depression and anxiety tend"},{"from":2505.42,"to":2506.59,"location":2,"content":"to be comorbid so"},{"from":2506.59,"to":2508.17,"location":2,"content":"people who have won often had the other"},{"from":2508.17,"to":2510.88,"location":2,"content":"so this points to the fact this points"},{"from":2510.88,"to":2512.44,"location":2,"content":"to the idea that perhaps there's some"},{"from":2512.44,"to":2514.36,"location":2,"content":"underlying representation that is"},{"from":2514.36,"to":2515.98,"location":2,"content":"similar across them that can be"},{"from":2515.98,"to":2518.2,"location":2,"content":"leveraged in a deep learning model with"},{"from":2518.2,"to":2521.5,"location":2,"content":"individual heads further specifying each"},{"from":2521.5,"to":2525.76,"location":2,"content":"of the different kinds of conditions and"},{"from":2525.76,"to":2527.65,"location":2,"content":"so what we found was that as we moved"},{"from":2527.65,"to":2529.99,"location":2,"content":"from logistic regression to the single"},{"from":2529.99,"to":2531.55,"location":2,"content":"task deep learning to the multi task"},{"from":2531.55,"to":2532.87,"location":2,"content":"deep learning we were able to get"},{"from":2532.87,"to":2534.94,"location":2,"content":"significantly better results and this"},{"from":2534.94,"to":2536.95,"location":2,"content":"was true both in the suicide risk case"},{"from":2536.95,"to":2539.47,"location":2,"content":"where we had a lot of data as well as"},{"from":2539.47,"to":2541.18,"location":2,"content":"the post traumatic stress disorder case"},{"from":2541.18,"to":2543.49,"location":2,"content":"where we had very little data the"},{"from":2543.49,"to":2545.23,"location":2,"content":"behavior here was a little bit different"},{"from":2545.23,"to":2548.29,"location":2,"content":"so going from logistic regression to"},{"from":2548.29,"to":2551.68,"location":2,"content":"single task deep learning when we had a"},{"from":2551.68,"to":2555.07,"location":2,"content":"lot of data as we did with the suicide"},{"from":2555.07,"to":2557.71,"location":2,"content":"risk had the single task deep learning"},{"from":2557.71,"to":2559.63,"location":2,"content":"model working better than the logistic"},{"from":2559.63,"to":2561.7,"location":2,"content":"regression model but when we had very"},{"from":2561.7,"to":2563.89,"location":2,"content":"few instances this is where the deep"},{"from":2563.89,"to":2565.48,"location":2,"content":"learning models really struggled a lot"},{"from":2565.48,"to":2568.75,"location":2,"content":"more and so the logistic regression"},{"from":2568.75,"to":2570.55,"location":2,"content":"models were actually much better but"},{"from":2570.55,"to":2572.11,"location":2,"content":"once we started adding heads for the"},{"from":2572.11,"to":2573.76,"location":2,"content":"comorbid different kinds of conditions"},{"from":2573.76,"to":2576.04,"location":2,"content":"the different kinds of tasks and that"},{"from":2576.04,"to":2577.93,"location":2,"content":"related to you know whether or not the"},{"from":2577.93,"to":2579.46,"location":2,"content":"person might be committing suicide and"},{"from":2579.46,"to":2582.1,"location":2,"content":"we were able to bump the accuracy way"},{"from":2582.1,"to":2586.21,"location":2,"content":"back up again and you know it's roughly"},{"from":2586.21,"to":2588.31,"location":2,"content":"120 at-risk individuals that we were"},{"from":2588.31,"to":2590.41,"location":2,"content":"able to collect in the suicide case that"},{"from":2590.41,"to":2591.76,"location":2,"content":"we wouldn't have otherwise been able to"},{"from":2591.76,"to":2597.94,"location":2,"content":"to notice as being at risk one of the"},{"from":2597.94,"to":2599.86,"location":2,"content":"approaches we took in this was to"},{"from":2599.86,"to":2601.87,"location":2,"content":"contextualize and consider the ethical"},{"from":2601.87,"to":2603.82,"location":2,"content":"dimensions of releasing this kind of"},{"from":2603.82,"to":2607.03,"location":2,"content":"technology so it's really common in NLP"},{"from":2607.03,"to":2609.82,"location":2,"content":"papers to give examples but this was an"},{"from":2609.82,"to":2611.29,"location":2,"content":"area where we decided that giving"},{"from":2611.29,"to":2613.15,"location":2,"content":"examples of like depressed language"},{"from":2613.15,"to":2615.43,"location":2,"content":"could be used to discriminate against"},{"from":2615.43,"to":2618.01,"location":2,"content":"people like add you know job interviews"},{"from":2618.01,"to":2619.93,"location":2,"content":"or something like that you know the sort"},{"from":2619.93,"to":2622.54,"location":2,"content":"of armchair psychology approach so we"},{"from":2622.54,"to":2624.1,"location":2,"content":"decided that while it was important to"},{"from":2624.1,"to":2625.99,"location":2,"content":"talk about the technique and the utility"},{"from":2625.99,"to":2627.64,"location":2,"content":"of multitask learning in a clinical"},{"from":2627.64,"to":2630.22,"location":2,"content":"domain and for bringing an inclusion of"},{"from":2630.22,"to":2632.47,"location":2,"content":"underrepresented subgroups it had to be"},{"from":2632.47,"to":2633.85,"location":2,"content":"balanced with the fact that there was a"},{"from":2633.85,"to":2636.82,"location":2,"content":"lot of risk in talking about depression"},{"from":2636.82,"to":2638.68,"location":2,"content":"and anxiety and how those kinds of"},{"from":2638.68,"to":2639.91,"location":2,"content":"things could be predicted"},{"from":2639.91,"to":2642.1,"location":2,"content":"so we tried to take a more balanced"},{"from":2642.1,"to":2644.35,"location":2,"content":"approach here and since then I've been"},{"from":2644.35,"to":2646.21,"location":2,"content":"putting ethical considerations in all of"},{"from":2646.21,"to":2648.4,"location":2,"content":"my papers it's becoming more and more"},{"from":2648.4,"to":2653.2,"location":2,"content":"common actually so another kind of"},{"from":2653.2,"to":2655.06,"location":2,"content":"approach that's now turning this on its"},{"from":2655.06,"to":2657.37,"location":2,"content":"head or you're trying to remove some"},{"from":2657.37,"to":2660.24,"location":2,"content":"effect mitigate bias in some way is"},{"from":2660.24,"to":2662.8,"location":2,"content":"adversarial multitask learning so I just"},{"from":2662.8,"to":2664.12,"location":2,"content":"talked about multitask learning now"},{"from":2664.12,"to":2665.64,"location":2,"content":"let's talk about the adversarial case"},{"from":2665.64,"to":2668.68,"location":2,"content":"and the idea in the adversarial case is"},{"from":2668.68,"to":2671.11,"location":2,"content":"that you have a few heads and one is"},{"from":2671.11,"to":2672.85,"location":2,"content":"predicting the main task and the other"},{"from":2672.85,"to":2674.53,"location":2,"content":"one is predicting the thing that you"},{"from":2674.53,"to":2677.38,"location":2,"content":"don't want to be affecting your models"},{"from":2677.38,"to":2679.69,"location":2,"content":"predictions so for example something"},{"from":2679.69,"to":2681.37,"location":2,"content":"like whether or not someone should be"},{"from":2681.37,"to":2683.56,"location":2,"content":"promoted based on you know their"},{"from":2683.56,"to":2685.84,"location":2,"content":"performance reviews and things like that"},{"from":2685.84,"to":2688.57,"location":2,"content":"and you don't want that to be affected"},{"from":2688.57,"to":2690.76,"location":2,"content":"by those gender ideally gender is"},{"from":2690.76,"to":2693.1,"location":2,"content":"independent of a promotion decision and"},{"from":2693.1,"to":2696.25,"location":2,"content":"so you can you can create a model for"},{"from":2696.25,"to":2698.17,"location":2,"content":"this that actually puts that"},{"from":2698.17,"to":2701.41,"location":2,"content":"independence criteria in place by saying"},{"from":2701.41,"to":2704.8,"location":2,"content":"I want to minimize my loss on the"},{"from":2704.8,"to":2707.05,"location":2,"content":"promotion while maximizing my loss on"},{"from":2707.05,"to":2709.03,"location":2,"content":"the gender and so how we're doing that"},{"from":2709.03,"to":2710.65,"location":2,"content":"is just predicting gender and then"},{"from":2710.65,"to":2713.44,"location":2,"content":"negating the gradient so removing the"},{"from":2713.44,"to":2716.53,"location":2,"content":"effect of that single it's this is"},{"from":2716.53,"to":2718.45,"location":2,"content":"another adversary approach so you might"},{"from":2718.45,"to":2720.04,"location":2,"content":"have been familiar with like generative"},{"from":2720.04,"to":2721.96,"location":2,"content":"adversarial networks so this is like two"},{"from":2721.96,"to":2724.62,"location":2,"content":"discriminators two different task heads"},{"from":2724.62,"to":2727.27,"location":2,"content":"where one is trying to do the task that"},{"from":2727.27,"to":2729.13,"location":2,"content":"we care about and the other one is"},{"from":2729.13,"to":2731.47,"location":2,"content":"removing the signal that we really don't"},{"from":2731.47,"to":2734.62,"location":2,"content":"want to be coming into play in our"},{"from":2734.62,"to":2737.02,"location":2,"content":"downstream predictions so this is a way"},{"from":2737.02,"to":2739.75,"location":2,"content":"of kind of putting this into practice so"},{"from":2739.75,"to":2741.97,"location":2,"content":"the probability of your output predicted"},{"from":2741.97,"to":2744.31,"location":2,"content":"output given the ground truth and you're"},{"from":2744.31,"to":2747.16,"location":2,"content":"sensitive attribute like gender is equal"},{"from":2747.16,"to":2749.29,"location":2,"content":"across all the different sensitive"},{"from":2749.29,"to":2750.7,"location":2,"content":"attributes or equal across all the"},{"from":2750.7,"to":2753.76,"location":2,"content":"different genders and that's an example"},{"from":2753.76,"to":2755.41,"location":2,"content":"of equality of opportunity and"},{"from":2755.41,"to":2757,"location":2,"content":"supervised learning being put into"},{"from":2757,"to":2758.92,"location":2,"content":"practice so this is one of the key"},{"from":2758.92,"to":2761.31,"location":2,"content":"fairness definitions it's equivalent to"},{"from":2761.31,"to":2763.78,"location":2,"content":"equal recall across different subgroups"},{"from":2763.78,"to":2766.15,"location":2,"content":"as I mentioned earlier and that's a"},{"from":2766.15,"to":2768.46,"location":2,"content":"model that will actually implement that"},{"from":2768.46,"to":2770.89,"location":2,"content":"or help you achieve that where you're"},{"from":2770.89,"to":2772.21,"location":2,"content":"saying that a classifiers output"},{"from":2772.21,"to":2773.62,"location":2,"content":"decision should be the same"},{"from":2773.62,"to":2776.32,"location":2,"content":"sensitive characteristics given what the"},{"from":2776.32,"to":2781.42,"location":2,"content":"what the correct decision should be okay"},{"from":2781.42,"to":2787.96,"location":2,"content":"so how are we on time there any"},{"from":2787.96,"to":2788.74,"location":2,"content":"questions so far"},{"from":2788.74,"to":2792.52,"location":2,"content":"we good okay cool so I'm gonna go into a"},{"from":2792.52,"to":2793.57,"location":2,"content":"little bit of a case study now and"},{"from":2793.57,"to":2796.6,"location":2,"content":"end-to-end system that Google has been"},{"from":2796.6,"to":2798.37,"location":2,"content":"working on my colleagues have been"},{"from":2798.37,"to":2800.83,"location":2,"content":"working on that is an NLP domain and"},{"from":2800.83,"to":2804.43,"location":2,"content":"deals with some of these bias issues so"},{"from":2804.43,"to":2806.77,"location":2,"content":"you can find out more about this work in"},{"from":2806.77,"to":2810.01,"location":2,"content":"papers at AI a s and 2018 and fat star"},{"from":2810.01,"to":2812.98,"location":2,"content":"tutorial 2019 called measuring and"},{"from":2812.98,"to":2814.72,"location":2,"content":"mitigating unintended bias and text"},{"from":2814.72,"to":2818.14,"location":2,"content":"classification and this came out of"},{"from":2818.14,"to":2821.89,"location":2,"content":"conversation AI which is a which is a"},{"from":2821.89,"to":2825.82,"location":2,"content":"product that's like it's part of this"},{"from":2825.82,"to":2828.19,"location":2,"content":"it's called a bet at Google it's a kind"},{"from":2828.19,"to":2830.74,"location":2,"content":"of spin-off company called jigsaw that"},{"from":2830.74,"to":2833.44,"location":2,"content":"focuses on trying to like combat abuse"},{"from":2833.44,"to":2836.83,"location":2,"content":"online and the conversation AI team is"},{"from":2836.83,"to":2838.27,"location":2,"content":"trying to use deep learning to improve"},{"from":2838.27,"to":2842.38,"location":2,"content":"online conversations and collaborate"},{"from":2842.38,"to":2844.27,"location":2,"content":"with a ton of different different people"},{"from":2844.27,"to":2847.93,"location":2,"content":"to do that so how this works is oh you"},{"from":2847.93,"to":2849.7,"location":2,"content":"can try it out to on perspective API"},{"from":2849.7,"to":2852.61,"location":2,"content":"comm so given some phrase like you're a"},{"from":2852.61,"to":2855.67,"location":2,"content":"dork it puts out a toxicity score"},{"from":2855.67,"to":2860.7,"location":2,"content":"associated to that like point nine one"},{"from":2860.7,"to":2863.23,"location":2,"content":"and the model starts sort of falsely"},{"from":2863.23,"to":2865.06,"location":2,"content":"associating frequently attacked"},{"from":2865.06,"to":2867.91,"location":2,"content":"identities with toxicity so this is a"},{"from":2867.91,"to":2871,"location":2,"content":"kind of false positive bias so I'm a"},{"from":2871,"to":2873.88,"location":2,"content":"proud tall person gets a model toxicity"},{"from":2873.88,"to":2877.78,"location":2,"content":"score of 0.18 I'm a proud gay person"},{"from":2877.78,"to":2880.81,"location":2,"content":"gets a toxicity model score of 0.69 and"},{"from":2880.81,"to":2883.84,"location":2,"content":"this is because these the term DEA tends"},{"from":2883.84,"to":2886.54,"location":2,"content":"to be used in really toxic situations"},{"from":2886.54,"to":2888.97,"location":2,"content":"and so the model starts to learn that"},{"from":2888.97,"to":2891.49,"location":2,"content":"gay itself is toxic but that's not"},{"from":2891.49,"to":2893.2,"location":2,"content":"actually what we want and we don't want"},{"from":2893.2,"to":2894.82,"location":2,"content":"these kinds of predictions coming out of"},{"from":2894.82,"to":2900.61,"location":2,"content":"the model so the bias is largely caused"},{"from":2900.61,"to":2903.01,"location":2,"content":"here by the data set imbalance again"},{"from":2903.01,"to":2904.75,"location":2,"content":"this is data kind of coming and running"},{"from":2904.75,"to":2905.62,"location":2,"content":"it's had again"},{"from":2905.62,"to":2908.62,"location":2,"content":"and so frequently attacks identities are"},{"from":2908.62,"to":2910.72,"location":2,"content":"really over-represented in toxic"},{"from":2910.72,"to":2912.22,"location":2,"content":"comments there's a lot of toxicity"},{"from":2912.22,"to":2915.52,"location":2,"content":"towards lbgtq identities it's really"},{"from":2915.52,"to":2917.17,"location":2,"content":"horrible to work on this stuff that like"},{"from":2917.17,"to":2921.18,"location":2,"content":"really can really affect you personally"},{"from":2921.18,"to":2924.34,"location":2,"content":"and one of the approaches that the team"},{"from":2924.34,"to":2927.31,"location":2,"content":"took was just to add non-toxic data from"},{"from":2927.31,"to":2930.04,"location":2,"content":"Wikipedia so helping to helping the"},{"from":2930.04,"to":2932.38,"location":2,"content":"model to understand that these kinds of"},{"from":2932.38,"to":2935.26,"location":2,"content":"terms can be used in you know more"},{"from":2935.26,"to":2942.07,"location":2,"content":"positive sorts of contexts one of the"},{"from":2942.07,"to":2944.62,"location":2,"content":"challenges with measuring how well the"},{"from":2944.62,"to":2946.24,"location":2,"content":"system was doing is that there's not a"},{"from":2946.24,"to":2949.63,"location":2,"content":"really nice way to have controlled"},{"from":2949.63,"to":2952.78,"location":2,"content":"toxicity evaluation so in real world"},{"from":2952.78,"to":2954.73,"location":2,"content":"conversation it can be kind of anyone's"},{"from":2954.73,"to":2958,"location":2,"content":"guess what the toxicity is of a specific"},{"from":2958,"to":2960.16,"location":2,"content":"sentence if you really want to control"},{"from":2960.16,"to":2962.46,"location":2,"content":"for different kind of subgroups or"},{"from":2962.46,"to":2964.39,"location":2,"content":"intersectional subgroups and it can be"},{"from":2964.39,"to":2967.06,"location":2,"content":"even harder to get real good data to"},{"from":2967.06,"to":2969.61,"location":2,"content":"evaluate properly so what the team ended"},{"from":2969.61,"to":2971.53,"location":2,"content":"up doing was developing a synthetic data"},{"from":2971.53,"to":2974.02,"location":2,"content":"approach so this is kind of like a bias"},{"from":2974.02,"to":2976.63,"location":2,"content":"mad libs where you take template"},{"from":2976.63,"to":2978.79,"location":2,"content":"sentences and you use those for"},{"from":2978.79,"to":2981.79,"location":2,"content":"evaluation this is a kind of evaluation"},{"from":2981.79,"to":2983.98,"location":2,"content":"you'd want to use in addition to your"},{"from":2983.98,"to":2987.88,"location":2,"content":"target downstream kind of data set but"},{"from":2987.88,"to":2990.37,"location":2,"content":"this helps you get at the biases"},{"from":2990.37,"to":2993.64,"location":2,"content":"specifically so some template phrase"},{"from":2993.64,"to":2996.13,"location":2,"content":"like I'm a proud blank person and then"},{"from":2996.13,"to":2998.02,"location":2,"content":"filling in different subgroup identities"},{"from":2998.02,"to":3000.45,"location":2,"content":"and you don't want to release a model"},{"from":3000.45,"to":3002.46,"location":2,"content":"unless you see that the scores across"},{"from":3002.46,"to":3005.61,"location":2,"content":"these different kinds of these different"},{"from":3005.61,"to":3007.08,"location":2,"content":"kinds of template sentences with"},{"from":3007.08,"to":3009.72,"location":2,"content":"synthetic the synthetic template"},{"from":3009.72,"to":3012.3,"location":2,"content":"sentences are relatively kind of the"},{"from":3012.3,"to":3015.09,"location":2,"content":"same across yeah all of the different"},{"from":3015.09,"to":3021.57,"location":2,"content":"model runs cool so some assumptions that"},{"from":3021.57,"to":3025.19,"location":2,"content":"they made in this was that the data set"},{"from":3025.19,"to":3027.75,"location":2,"content":"didn't have annotated bias and they"},{"from":3027.75,"to":3029.43,"location":2,"content":"didn't do any causal analysis because"},{"from":3029.43,"to":3030.48,"location":2,"content":"they were just trying to focus in"},{"from":3030.48,"to":3034.76,"location":2,"content":"particular on this toxicity problem and"},{"from":3034.76,"to":3038.88,"location":2,"content":"they used a CNN convolutional yeah you"},{"from":3038.88,"to":3040.14,"location":2,"content":"guys know"},{"from":3040.14,"to":3041.73,"location":2,"content":"with pre-trained glove embeddings this"},{"from":3041.73,"to":3042.84,"location":2,"content":"is probably like your bread and butter"},{"from":3042.84,"to":3044.46,"location":2,"content":"fruit and gloves embeddings I'm sure you"},{"from":3044.46,"to":3046.14,"location":2,"content":"know all about this and we're Tyvek cool"},{"from":3046.14,"to":3050.24,"location":2,"content":"curious implementation of this and and"},{"from":3050.24,"to":3052.5,"location":2,"content":"using these kind of data augmentation"},{"from":3052.5,"to":3055.71,"location":2,"content":"approaches both a Wikipedia kind of"},{"from":3055.71,"to":3057.9,"location":2,"content":"approach as well as actually collecting"},{"from":3057.9,"to":3060.6,"location":2,"content":"positive statements about LGBTQ identity"},{"from":3060.6,"to":3062.58,"location":2,"content":"so there's this project called Project"},{"from":3062.58,"to":3065.1,"location":2,"content":"respected Google where we go out and and"},{"from":3065.1,"to":3067.14,"location":2,"content":"talk to to people who identify as queer"},{"from":3067.14,"to":3069,"location":2,"content":"or people who have friends who do and"},{"from":3069,"to":3070.98,"location":2,"content":"like talk about this in a positive way"},{"from":3070.98,"to":3073.83,"location":2,"content":"and we add this as data so we can"},{"from":3073.83,"to":3075.66,"location":2,"content":"actually know that this is can be a"},{"from":3075.66,"to":3079.47,"location":2,"content":"positive thing and in order to measure"},{"from":3079.47,"to":3082.08,"location":2,"content":"the model performance here again it's"},{"from":3082.08,"to":3083.76,"location":2,"content":"looking at the differences across"},{"from":3083.76,"to":3085.47,"location":2,"content":"different subgroups and trying to"},{"from":3085.47,"to":3087.78,"location":2,"content":"compare also the subgroup performance to"},{"from":3087.78,"to":3089.94,"location":2,"content":"some sort of general distribution so"},{"from":3089.94,"to":3092.73,"location":2,"content":"here they use a UC where a UC is"},{"from":3092.73,"to":3094.44,"location":2,"content":"essentially the probability that a model"},{"from":3094.44,"to":3097.17,"location":2,"content":"will give a randomly selected positive"},{"from":3097.17,"to":3099.66,"location":2,"content":"example a higher score than a randomly"},{"from":3099.66,"to":3103.23,"location":2,"content":"selected a negative example so here you"},{"from":3103.23,"to":3105.03,"location":2,"content":"can see some toxic comments and non"},{"from":3105.03,"to":3107.52,"location":2,"content":"toxic comments with that example sort of"},{"from":3107.52,"to":3112.35,"location":2,"content":"low a you see here this is a example"},{"from":3112.35,"to":3114.57,"location":2,"content":"with a high AUC so the model is doing a"},{"from":3114.57,"to":3116.25,"location":2,"content":"relatively good job separating these two"},{"from":3116.25,"to":3119.94,"location":2,"content":"kinds of comments and there are"},{"from":3119.94,"to":3121.23,"location":2,"content":"different kinds of biases that they've"},{"from":3121.23,"to":3124.02,"location":2,"content":"defined in this work so low subgroup"},{"from":3124.02,"to":3125.46,"location":2,"content":"performance means that the model"},{"from":3125.46,"to":3127.38,"location":2,"content":"performs worse on subgroup comments than"},{"from":3127.38,"to":3129.93,"location":2,"content":"it does on comments overall and the"},{"from":3129.93,"to":3131.31,"location":2,"content":"metric they've introduced to measure"},{"from":3131.31,"to":3134.84,"location":2,"content":"this is called subgroup a you see"},{"from":3134.84,"to":3137.67,"location":2,"content":"another one is subgroup shift and that's"},{"from":3137.67,"to":3139.17,"location":2,"content":"when the model systematically scores"},{"from":3139.17,"to":3142.68,"location":2,"content":"comments from some subgroup higher so"},{"from":3142.68,"to":3144.87,"location":2,"content":"this is sort of like to the right and"},{"from":3144.87,"to":3147.3,"location":2,"content":"then there's also this background"},{"from":3147.3,"to":3149.55,"location":2,"content":"positive subgroup negatives shifting to"},{"from":3149.55,"to":3156.33,"location":2,"content":"the left yeah yeah this sort of saying"},{"from":3156.33,"to":3157.77,"location":2,"content":"when I said it can go either way to the"},{"from":3157.77,"to":3159.12,"location":2,"content":"right or the left and there's just kind"},{"from":3159.12,"to":3160.95,"location":2,"content":"of different metrics that can define"},{"from":3160.95,"to":3166.87,"location":2,"content":"each of these and the results in this"},{"from":3166.87,"to":3168.88,"location":2,"content":"sort of going through not only just"},{"from":3168.88,"to":3171.01,"location":2,"content":"looking at you know qualitative examples"},{"from":3171.01,"to":3173.74,"location":2,"content":"and general evaluation metrics but also"},{"from":3173.74,"to":3175.48,"location":2,"content":"focusing in on some of the key metrics"},{"from":3175.48,"to":3177.31,"location":2,"content":"defined for this work these sort of AUC"},{"from":3177.31,"to":3179.32,"location":2,"content":"based approaches and they were able to"},{"from":3179.32,"to":3181.09,"location":2,"content":"see significant differences in the"},{"from":3181.09,"to":3183.22,"location":2,"content":"original release which didn't account"},{"from":3183.22,"to":3185.11,"location":2,"content":"for any of these unintended biases and"},{"from":3185.11,"to":3187.63,"location":2,"content":"downstream releases which did which"},{"from":3187.63,"to":3189.85,"location":2,"content":"incorporated this kind of normative data"},{"from":3189.85,"to":3192.01,"location":2,"content":"that said the sort of things that we"},{"from":3192.01,"to":3195.63,"location":2,"content":"thought the model should be learning"},{"from":3195.63,"to":3199.36,"location":2,"content":"cool so the last thing to keep in mind"},{"from":3199.36,"to":3201.58,"location":2,"content":"as you sort of develop and work towards"},{"from":3201.58,"to":3204.43,"location":2,"content":"a creating deeper better models is to"},{"from":3204.43,"to":3207.46,"location":2,"content":"release responsibly so this is a project"},{"from":3207.46,"to":3208.72,"location":2,"content":"I've been working on with a ton of"},{"from":3208.72,"to":3210.49,"location":2,"content":"different people called model cards for"},{"from":3210.49,"to":3212.89,"location":2,"content":"model reporting it's a it's a little bit"},{"from":3212.89,"to":3215.17,"location":2,"content":"of like the next step after data sheets"},{"from":3215.17,"to":3218.47,"location":2,"content":"for datasets where data sheets for"},{"from":3218.47,"to":3220.63,"location":2,"content":"datasets focus on information about the"},{"from":3220.63,"to":3223.27,"location":2,"content":"data model cards for model reporting"},{"from":3223.27,"to":3225.27,"location":2,"content":"focuses on information about the model"},{"from":3225.27,"to":3228.52,"location":2,"content":"so it captures what it does how it works"},{"from":3228.52,"to":3232.27,"location":2,"content":"why it matters and one of the key ideas"},{"from":3232.27,"to":3234.52,"location":2,"content":"here is disaggregated and intersectional"},{"from":3234.52,"to":3237.85,"location":2,"content":"evaluation so it's not enough anymore"},{"from":3237.85,"to":3239.62,"location":2,"content":"to put out human centered technology"},{"from":3239.62,"to":3242.68,"location":2,"content":"that just has some vague overall score"},{"from":3242.68,"to":3244.12,"location":2,"content":"associated to it"},{"from":3244.12,"to":3245.71,"location":2,"content":"you actually need to understand how it"},{"from":3245.71,"to":3247.06,"location":2,"content":"works across different subpopulations"},{"from":3247.06,"to":3249.49,"location":2,"content":"and you have to understand what the data"},{"from":3249.49,"to":3252.94,"location":2,"content":"is telling you that um so here's some"},{"from":3252.94,"to":3255.22,"location":2,"content":"example details that model card would"},{"from":3255.22,"to":3257.65,"location":2,"content":"have who it's developed by what the"},{"from":3257.65,"to":3259.93,"location":2,"content":"intended uses so that it doesn't start"},{"from":3259.93,"to":3261.4,"location":2,"content":"being used in ways that it's not"},{"from":3261.4,"to":3264.04,"location":2,"content":"intended to be used the factors that are"},{"from":3264.04,"to":3265.87,"location":2,"content":"likely to be affected by"},{"from":3265.87,"to":3267.55,"location":2,"content":"disproportionate performance of the"},{"from":3267.55,"to":3270.04,"location":2,"content":"model so different kinds of identity"},{"from":3270.04,"to":3273.16,"location":2,"content":"groups things like that the metrics that"},{"from":3273.16,"to":3275.41,"location":2,"content":"that you're deciding to use in order to"},{"from":3275.41,"to":3277.12,"location":2,"content":"understand the fairness of the model or"},{"from":3277.12,"to":3279.43,"location":2,"content":"the different performance of the model"},{"from":3279.43,"to":3281.08,"location":2,"content":"across different kinds of subgroups and"},{"from":3281.08,"to":3283.66,"location":2,"content":"factors information about the evaluation"},{"from":3283.66,"to":3287.23,"location":2,"content":"data and training data as well as"},{"from":3287.23,"to":3289.9,"location":2,"content":"ethical considerations so what were some"},{"from":3289.9,"to":3291.64,"location":2,"content":"of the things you took into account or"},{"from":3291.64,"to":3293.55,"location":2,"content":"what are some of the risks and benefits"},{"from":3293.55,"to":3297.25,"location":2,"content":"that that are relevant to this model and"},{"from":3297.25,"to":3299.2,"location":2,"content":"additional caveats and recommendations"},{"from":3299.2,"to":3300.46,"location":2,"content":"so for example"},{"from":3300.46,"to":3302.89,"location":2,"content":"in the conversation the eye case they're"},{"from":3302.89,"to":3304.72,"location":2,"content":"working with synthetic data so this is"},{"from":3304.72,"to":3306.64,"location":2,"content":"the sort of limitation of the evaluation"},{"from":3306.64,"to":3308.95,"location":2,"content":"that's important to understand because"},{"from":3308.95,"to":3310.87,"location":2,"content":"it can tell you a lot about the biases"},{"from":3310.87,"to":3312.49,"location":2,"content":"but doesn't tell you a lot about how it"},{"from":3312.49,"to":3318.01,"location":2,"content":"works generally and then the key"},{"from":3318.01,"to":3320.56,"location":2,"content":"component in the quantitative section of"},{"from":3320.56,"to":3322.06,"location":2,"content":"the model card is to have this both"},{"from":3322.06,"to":3323.74,"location":2,"content":"intersectional and disaggregated"},{"from":3323.74,"to":3325.81,"location":2,"content":"evaluation and from here you trivially"},{"from":3325.81,"to":3327.34,"location":2,"content":"get two different kinds of fairness"},{"from":3327.34,"to":3329.8,"location":2,"content":"definitions the closer you get to parity"},{"from":3329.8,"to":3331.57,"location":2,"content":"across subgroups the closer you're"},{"from":3331.57,"to":3332.71,"location":2,"content":"getting to something that is"},{"from":3332.71,"to":3336.94,"location":2,"content":"mathematically fair okay so hopefully by"},{"from":3336.94,"to":3338.2,"location":2,"content":"paying attention to these kinds of"},{"from":3338.2,"to":3340.36,"location":2,"content":"approaches taking into account all these"},{"from":3340.36,"to":3341.8,"location":2,"content":"kinds of things we can move from"},{"from":3341.8,"to":3344.14,"location":2,"content":"majority representation of data in our"},{"from":3344.14,"to":3346.24,"location":2,"content":"models to something more like diverse"},{"from":3346.24,"to":3349.99,"location":2,"content":"representation for more ethical AI okay"},{"from":3349.99,"to":3350.77,"location":2,"content":"that's it"},{"from":3350.77,"to":3352.2,"location":2,"content":"Thanks"},{"from":3352.2,"to":3359.89,"location":2,"content":"[Applause]"}]}