2022-03-12 08:11:00 CAFIAC FIX 2022-01-17 08:11:00 CAFIAC FIX 2022-01-11 08:11:00 CAFIAC FIX 2021-12-27 08:20:00 CAFIAC FIX 2021-12-07 23:12:00 @ShamKakade6 @ChanZuckerberg @Harvard @blsabatini Congratulations! This is huge. 2021-12-01 21:35:23 @danupon @thesasho Congratulations! 2021-11-09 19:24:18 @KordingLab @beenwrekt Would love to learn more about what you have in mind. DM appreciated! Open to changing my mind about ND. 2021-11-09 19:23:58 @KordingLab @beenwrekt No strong opinion, but I personally prefer ND in order to preserve the intellectual perspective and narrative of the book as a whole. I'd prefer to grant reasonable exceptions rather than agreeing to any conceivable remix/alteration up front. 2021-11-06 23:20:00 CAFIAC FIX 2021-11-01 19:20:00 CAFIAC FIX 2021-11-01 17:30:00 CAFIAC FIX 2021-08-19 18:00:22 RT @thejonullman: Last Spring, DP co-inventor Adam Smith and I jointly taught an intro course on DP. We've finally gotten around to making… 2021-05-17 15:00:06 @geomblog @WHOSTP @AlondraNelson46 @BrownUniversity @BrownCSDept @Brown_DSI @senykam @UtahSoC Congratulations and best wishes! 2021-05-17 15:00:06 @geomblog @WHOSTP @AlondraNelson46 @BrownUniversity @BrownCSDept @Brown_DSI @senykam @UtahSoC Congratulations and best wishes! 2021-05-03 14:46:37 @jasondeanlee @paulwgoldberg My guess is that when epsilon is part of the input description it's log(1/epsilon) bits long. So, an efficient algorithm would have to have a log(1/epsilon) dependence. Correct me if I'm wrong. Nice result! 2021-03-22 20:36:29 RT @MD4SG: We're thrilled to announce the inaugural ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EA… 2021-03-16 16:19:07 @s_scardapane @beenwrekt Paper link: https://t.co/XaPQdxQrjD by @beenwrekt @BeccaRoelofs @lschmidt3 @Vaishaal 2021-03-12 16:07:58 @s_scardapane @permutans @beenwrekt Yes, you may. There will also be a published print edition in the future. Thanks for your interest. 2021-02-14 17:48:55 RT @mona_sloane: YES -- I have been waiting for this! "PATTERNS, PREDICTIONS, AND ACTIONS - A story about machine learning" will be a fanta… 2021-02-12 17:28:11 @docmilanfar @beenwrekt @UCBerkeley No better theft! And we were so lucky to have our students help us. 2020-12-12 16:53:52 RT @angelamczhou: Happening now! Very exciting speaker lineup: @red_abebe, @hamsabastani, Yiling Chen, @mrtz, Mitsue Iwata, @maxkasy, and… 2020-12-03 04:40:06 @timnitGebru @JeffDean Timnit, you are a bold thinker and a role model for our community. Your future in research is bright and the community will be better for it. I am gravely concerned about what Google has done here. 2020-11-24 15:39:43 The new author fees at journals like Nature are a play against collective action. Publishers recognize that authors are less likely to organize and negotiate as a group than universities. The best response for authors is therefore to collectively engage or abandon these journals. 2020-11-19 04:13:14 @JSEllenberg @beenwrekt It's inspired by matrix rigidity over GF(2). 2020-11-19 04:12:28 @JSEllenberg @beenwrekt You might like this: https://t.co/NY17vNw340 2020-11-19 01:33:26 @thegautamkamath Man, I'm ready to hike the Pacific Crest Trail. 2020-10-07 16:33:08 RT @red_abebe: We're going into our 4th round of the @black_in_ai grad app program and costs associated with GRE/TOEFL + apps are *the main… 2020-10-02 16:59:44 RT @Andrea__M: If you are interested doing research on the mathematics of machine learning, consider applying for a postdoc with our collab… 2020-09-29 17:21:33 RT @boazbaraktcs: Thomas Vidick has been a role model in terms of scientific results, a role model in terms of advising students, and is no… 2020-09-25 20:18:14 This article by @uhlily illuminates serious problems with causal inference on racial discrimination, the trouble with "direct effects of race", and the harm in reducing substantive normative debate to purely statistical routine. A rare must-read. https://t.co/o9sszfOjmd 2020-09-01 15:13:18 @k__niki @helmholtz_ai @HelmholtzMunich Congratulations, Niki! I look forward to the great things your group is going to do! 2020-08-24 14:34:21 @boazbaraktcs @valeriestrauss @washingtonpost A well reasoned and persuasive argument! Perhaps this 8th grader can help her dad with the next paper intro! 2020-08-21 16:51:44 RT @SophieHeloise: I spent the weekend reading Ofqual's technical report explaining their grading algorithm (so you don't have to). Here is… 2020-08-14 02:12:14 @natfriedman You will enjoy this article by @kevinbaker about it: https://t.co/0Xu2MEqV8j 2020-06-23 22:31:18 My point is not to weigh in on the criminal justice debate, but to illustrate that questions of justice around predictive modeling don't have the kind of modularity that would support a meaningful discussion of data set bias in isolation. 2020-06-23 22:31:17 Apropos of current discussions, here's an example I found helpful in understanding why opting for prediction as a solution concept on its own (regardless of data and modeling choices) is already a consequential political act that deprioritizes alternatives. https://t.co/SW9ITvZKtU 2020-06-19 00:27:41 This is an awesome bike shop. My daughter is riding a beautiful bike the owner built for his daughter. https://t.co/4tRGdfq4r9 2020-05-19 14:28:53 RT @red_abebe: In a 15+ year old dream come true, I'm happy to share that I'm going to be a professor! I will be joining @Berkeley_EECS as… 2020-04-28 15:02:50 @cashewmake2 @SFarley_Charlie Yeah, I mean it's just a matplotlib bar plot with .xkcd() called. 2020-04-27 22:56:07 @SFarley_Charlie Sure thing. 2020-04-27 19:56:19 @SFarley_Charlie Yes. Opening slide from my 2017 class. It's traveled far and wide since. 2020-04-21 03:07:18 @forrestercole @stephan_gfx @CasualEffects @stevelavietes Very cool! I implemented similar mechanics for this floppy bird example in my optimization class: https://t.co/a9Nk0HvU5j 2020-04-16 01:47:37 @annejonas Rhododendron aureum? Another one of my favs after the dog wood you posted recently. 2020-04-14 20:25:34 @DKThomp @asociologist We call the latter performative prediction https://t.co/sr9cQqxGeN 2020-03-24 15:53:36 @annejonas Great catch. One of my favorites! The one in our front yard is just about to flower. 2020-03-20 19:47:16 There's a lot of good stuff under the hood, such as plug-in support for the OpenAI Gym reinforcement learning API, as well as a range of causal inference methods supported out of the box. 2020-03-20 19:47:15 Today we're launching WhyNot, an experimental sandbox for decision making in dynamics. WhyNot is a Python package connecting tools from causal inference and reinforcement learning with a range of complex simulators. Check it out: https://t.co/zXCIejXIK1 2020-03-07 15:35:56 Due to the coronavirus, we'll be moving this university to Twitch. Professors, please stock up on Mountain Dew and purchase the Secretlab OMEGA Stealth gaming chair with lumbar pillow. 2020-03-05 03:02:04 RT @TCS_plus: Sign up for next week's talk! Thomas Steinke (@shortstein) from @IBMResearch on "Reasoning About Generalization via Condition… 2020-03-03 20:04:02 @ccanonne_ Yes, strategic classification is a special case of performative prediction. Lots of work in that area. 2020-03-03 17:11:07 RT @stepheaneff: Also relevant for clinical predictive algorithms! When an algorithm suggests that a patient is high risk for a given diagn… 2020-03-03 16:37:54 RT @paulgp: Cool! This seems like a great topic. This is super salient for economists studying monetary policy - reminds me of the issues t… 2020-03-03 16:33:24 To be clear, this is obviously still far from applying to FiveThirtyEight directly, but it would be an interesting case study. 2020-03-03 16:18:37 Indeed, prediction is often performative: The prediction influences the outcome it's trying to predict. Classical theory neglects this aspect, but see this recent work: https://t.co/sr9cQqxGeN https://t.co/IbEKsN8Kg4 2020-03-02 19:10:16 RT @nberpubs: Ride-hailing's launch in US cities has led to increases in congestion and traffic fatalities that have persisted over time, f… 2020-03-02 17:18:39 @KLdivergence @Penn Congratulations! Wonderful news. 2020-02-25 00:42:39 @geomblog @Aaroth @thegautamkamath I see you and raise you to inviting the PC member who co-authored the paper. 2020-02-21 02:54:58 The role of AI in medicine aside, ethics rarely reduces to trolley problems like these. https://t.co/9GbH6nE3qx 2020-02-18 05:49:22 @Aaroth @RowRyRog @JinshuoD Nice! Does this also apply to the noisy max implementation or do you have to get the exact ExpMech distribution right? 2020-01-30 16:42:56 Petition to change the Google Slides default animation to a plain step effect, no delay, no fade-in. 2020-01-28 22:35:16 @shortstein Yeah, it's all over my feed. It's basically coronavirus, impeachment, and your paper. 2020-01-22 03:46:33 @BooleanAnalysis Fall 2020? You really live in the future. Who won the election? 2019-12-27 19:34:28 @KLdivergence Congratulations, Kristian! Another well deserved honor. 2019-12-27 19:17:17 @natefoster But I deleted Strava, so I'm no longer on that list either 2019-12-27 19:07:11 Thrilled to report that 2019 was the year I had successfully evaded all "under 35" lists to my knowledge. It's been a long road and I want to thank all who've supported me over the years. 2019-12-19 20:17:06 @thojansen Congrats! Very curious to hear what's next. 2019-12-19 18:39:56 @_hylandSL @thesasho You can see this in experiment: MNIST is close to a, say, 100-dimensional subspace. Take something that's very far from the MNIST subspace, say, an image of a car. It should be easy to tell from the output of SGD whether or not that image was in the data. 2019-12-19 18:37:33 @_hylandSL @thesasho Thanks for the clarification. Same issue in general though: Your points could be close to a lower dimensional subspace, add a point far from the subspace and you change the natural SGD distribution significantly. 2019-12-19 18:02:17 @thesasho @_hylandSL Plain SGD cannot be differentially private in the overparameterized regime even for linear models and squared loss: The model always lies in an n-dimensional subspace of R^d, changing a single data point changes the support of the output distribution of SGD. Hence, no DP. 2019-12-19 17:59:59 @_hylandSL I think the bigger issue is that even if the marginals of the SGD steps looked Gaussian, they'd be strongly correlated. 2019-12-19 17:56:16 @thesasho @_hylandSL I also stumbled over this. I think the crucial assumption is that the "noise added by SGD" is Gaussian. The DP claim holds for Gaussian noise with the same variance as the actual SGD distribution. You're right that SGD isn't DP, and leaks data via inversion attacks. 2019-12-13 15:51:51 RT @emilymbender: Ethical considerations are not a separate concern from whether the AI research is “sound”: sound AI research requires not… 2019-12-12 23:01:59 @alexhanna Unfortunately, this stuff comes straight from the top. Here's prior work by leaders in the field from MIT: https://t.co/eAXNAXzGl7 - the attempt at an "ethics" section misses the point. 2019-12-12 18:10:54 @geomblog Yes, I figured you knew those works. 2019-12-12 15:51:12 @random_walker The Friedman/Nissenbaum is certainly important early pioneering work in this area, but sorting out who was "first" will depend on your definitions. 2019-12-12 15:46:41 Fairness concerns were also raised about the residency match in the 50s already. See this paper by Al Roth on the history: https://t.co/bYT6vPeU4T HT: @random_walker The quote below is from 1950: https://t.co/Nfs5PUPF9a 2019-12-12 15:46:40 Recognition of the test bias along demographic lines probably goes back to the early 20th century. See the discussion around issues with the Army Alpha test. https://t.co/FaeVbsnhF6 https://t.co/nzeR6D2xAb 2019-12-12 15:46:39 The educational testing folks stumbled over issues of bias very early. Anne Cleary wrote pioneering papers about test bias in the 60s, and computers were used to process tests even earlier. https://t.co/YmIj7HZggd https://t.co/LDiu550RBS 2019-12-10 16:08:23 RT @sapiezynski: Showing liberal ads to conservatives (or cons ads to libs) can cost 3x more than showing an "aligned" ad to the same audie… 2019-12-10 15:46:24 RT @korolova: We show that Facebook’s ad delivery algorithms hinder political campaigns’ ability to reach diverse voters, impose different… 2019-12-04 20:21:41 RT @healthyalgo: My paper on Differential Privacy in the US Census (coauthored with Samantha Petti) is now available as a preprint: https:/… 2019-11-29 22:46:33 @QuentinHardt Yeah, I was kinda thinking the same when I saw this. 2019-11-28 20:49:19 Google Play Thanksgiving recommendations are a bit subtle. https://t.co/uzJ88s2IOw 2019-11-27 17:33:40 A new book chapter on testing discrimination in practice is now available at https://t.co/70pwvMkUf4 https://t.co/NuKYccTlBk 2019-11-27 02:32:29 If you want to learn more, there's an excellent monograph by Dwork and @Aaroth available online: https://t.co/6fKD0s9xs3 2019-11-27 02:16:08 I just gave two guest lectures on privacy-preserving data analysis and differential privacy in Berkeley's new data science class. Here are the slides if helpful: https://t.co/dISxExe7er 2019-11-26 21:10:42 RT @kevinbaker: Does anyone have any #histsci/#sts lit recommendations on support vector machines an/or Vladimir Vapnik? Not looking to sco… 2019-11-19 20:03:13 RT @amyxzh: Hi all! Please RT! I am recruiting PhD students @uwcse! My new lab will be reimagining online social and collaborative systems… 2019-11-06 19:23:01 @danbjork @seanjtaylor @alex_peys @jblumenstock Here's the link to a recent paper on this: https://t.co/acgbJuZU27 2019-11-04 16:38:30 @BachFrancis Great post! For those of you want to play around with this a bit more, here's a code book about this I did for my optimization class a while back: https://t.co/MnW02c9sqU 2019-11-01 16:09:35 RT @zephoria: This week, the @censusbureau released a set of demonstration files for those who work with census data to examine how differe… 2019-10-30 15:26:53 @autreche @SmithaMilli Yes, absolutely. We'll update the arxiv version shortly. Thank you for bringing this up. 2019-10-18 16:29:37 If this thread had you interested in overfitting in machine learning benchmarks, you can watch a longer talk here: https://t.co/KzQCaOEaKO (Slides here: https://t.co/7xI50yq7eI) https://t.co/iVWZ2M6kL7 2019-10-08 20:28:53 @BEBischof @osageev Topological quantum field theory, to be more precise. 2019-10-08 20:27:24 @Blizzard_Ent @EFF @Blizzard_Ent has denied my written request to delete my account. 2019-10-08 19:54:07 @lishali88 @chrisemoody @stitchfix Yes, I know they have a strong data science team. It's just the headline. 2019-10-08 19:05:26 Stich Fix is selling something called pants, a concept from quantum mechanics. https://t.co/laKu6KDJVJ 2019-10-08 15:59:22 @timnitGebru @Pinboard People in the east bay seem to hold significantly fewer stakes. 2019-10-08 15:20:03 .@Blizzard_Ent requires a government ID to close an account that never required an ID or its associated information to create in the first place. Why? What do they even compare the ID against? Do folks here know if this practice is allowed? What are my rights? CC @EFF 2019-10-03 23:49:09 @realnamepolicy It took me hours to get this joke. Kudos! 2019-10-03 23:24:06 @colinraffel Point taken. Will delete tweet. 2019-10-03 23:23:10 RT @colinraffel: @mrtz https://t.co/jR0LaXmfEh ? 2019-10-03 21:29:48 @yoavgo True. There are some concrete research directions after all. 2019-10-03 21:04:10 "Concrete research question" is an oxymoron. 2019-10-01 20:11:52 The decision in the important SFFA v Harvard admissions lawsuit has just been posted and from a very brief look it appears to be mostly a win for Harvard. https://t.co/qpXnlaCpFQ 2019-09-20 14:18:41 RT @Pinboard: Strive to live a life where you never have to hear someone say 'stakeholder' 2019-09-20 03:15:54 @Aaroth @mikekimbackward @zstevenwu @mkearnsupenn @LMiracchi The teacher to student ratio at Penn is impeccable. 2019-09-19 20:17:51 @IAmSamFin @s010n @random_walker Here it is now btw: https://t.co/u0tknvcw0A 2019-09-19 17:14:17 We just open sourced the production setup behind https://t.co/og0iNDglC0. Check it out here: https://t.co/u0tknvcw0A It's a simple and reusable setup that produces both html and pdf with a few useful features. We hope you'll find it useful for books, lecture notes, and projects. 2019-09-18 15:22:58 @fanaeepour The book focuses on decision making, but what is classification in a more narrow technical sense is only one of seven chapters. 2019-09-09 16:41:06 We're keeping track of contributors on this page https://t.co/Pk2svHV2kr. In the unfortunate event that you we forgot to include you, please remind us and accept our apology. 2019-09-09 16:39:01 PDF available here: https://t.co/DL2Pbkw4SZ 2019-09-09 16:39:00 We're excited to release a draft of the causality chapter at https://t.co/og0iNDglC0. There's been lots of interest in the topic and we hope the chapter is a helpful introduction. https://t.co/NuKYccTlBk 2019-08-29 15:12:24 @Aaroth The animated gif convinced me that this is a solid purchase decision. Done. 2019-08-27 03:17:37 Seen in Causation, Prediction, and Search by Spirtes, Glymour, Scheines. 2019-08-27 03:15:10 Interesting comment on an interesting thing George Box [almost] said. https://t.co/wMC0x493QJ 2019-08-27 00:59:11 RT @histoftech: Hey west coast frens: I’ll be giving a talk at @UCBerkeley the afternoon of September 19, at the Center for Science, Techno… 2019-08-16 04:19:30 RT @AcharyaJayadev: How much can one overfit on a multiclass classification test set with n examples, and m classes by using it k times? T… 2019-08-14 01:58:33 RT @DMulliganUCB: I am totally thrilled to share the news that Dr. Jennifer Chayes will be the first Associate Provost of the Division of D… 2019-07-31 17:50:18 RT @morgangames: So thrilled to be staying on at the @BerkeleyISchool more permanently - I can't think of a place I'd rather be < 2019-07-24 21:56:29 @likeavass @kly_cl See, I'm the other type. To me these are all generic looking red sneakers. 2019-07-24 19:08:12 HT @likeavass @kly_cl 2019-07-24 19:06:40 I saw this tweet while sitting next to @beenwrekt and so I looked at our shoes. https://t.co/J0019vkQsh 2019-07-16 16:58:12 RT @mer__edith: Yesterday was my last day at Google. It's been a long, wild road, and I am thrilled to be officially stepping into my role… 2019-07-10 02:18:11 @histoftech @tdietterich @ruchowdh @mer__edith @merbroussard @safiyanoble @ruha9 @ubiquity75 @gleemie Indeed, I only learned about this through your tweets @histoftech - so let me express my gratitude for bringing this to wider attention. Your points are all well taken. I'm a keen follower of the terrific scholars you mention. 2019-07-09 20:08:58 RT @histoftech: Weizenbaum was right, but there’s plenty of tech critics working *right now* on the current state of these problems & 2019-07-09 15:21:09 Source: https://t.co/oEddpL5DAd 2019-07-09 15:21:08 Weizenbaum wrote this in 1972 and it remains current in how it simultaneously describes today's techies and tech critics. https://t.co/N0Dl8D9Bc7 2019-07-08 17:15:23 RT @bwhitman: @kcimc i once heard a Coldplay song so bad i started a company 2019-06-26 02:18:31 RT @histoftech: all. of. this. https://t.co/6iIjMrG9ee 2019-06-20 21:00:16 @mraginsky exit() 2019-06-18 20:46:19 @jugander @Smily This has really been my most technical research contribution all year. 2019-06-13 05:08:23 @ericjschmidt @npseaver As a matter of fact, they do. This one seems to scale Africa more accurately relative to Europe: https://t.co/y5Y18F0PVN 2019-06-06 15:20:20 @prfsanjeevarora @goodfellow_ian @_beenkim @julius_adebayo Also, @jmgilmer and @michaelmuelly - turns out everyone's on Twitter :-) 2019-06-06 15:19:04 @prfsanjeevarora @goodfellow_ian @_beenkim Great! Main author on sanity checks paper is @julius_adebayo 2019-05-31 17:17:26 @SmithaMilli @Eric_Wallace_ @geomblog @RDBinns @ihsgnef @NicolasPapernot @ChristophMolnar @fatconference @lschmidt3 @ancadianadragan True, but if the explanation is supposed to reveal information about the population rather than the sample, differential privacy would not prevent you from doing so in principle. 2019-05-31 15:38:06 @Eric_Wallace_ @geomblog @RDBinns @ihsgnef @NicolasPapernot @ChristophMolnar @SmithaMilli @fatconference @lschmidt3 @ancadianadragan Indeed, in the case of model inversion from predictions, Dwork and Feldman proposed this remedy. It's worth thinking a bit about how well this defense would work for saliency maps. 2019-05-31 14:46:25 @geomblog @RDBinns @ihsgnef @NicolasPapernot @ChristophMolnar @SmithaMilli @fatconference Right, @ihsgnef had already cited it. It's with @SmithaMilli @lschmidt3 @ancadianadragan - would be fun to extend it to LIME 2019-05-31 14:26:17 @roydanroy Thanks. Curious where and what in that paper is the connection to #4? 2019-05-31 02:57:42 @guyi You make an important point. Whatever negative consequences there are to benchmark overuse they are not apparent from the narrow train vs test paradigm. 2019-05-30 18:44:49 @IAmSamFin @s010n @random_walker It's a custom pandoc setup with various tweaks and filters. It's not pretty unfortunately, but compiles reasonably fast. 2019-05-30 18:40:45 @IAmSamFin @s010n @random_walker We're planning on open sourcing the production setup once it's fully stable. Should be out later this year. The code for figures is already available here: https://t.co/QxTSwWaK1B 2019-05-30 18:04:45 What limits overfitting in the machine learning community are sociotechnical factors, hard to capture from a purely statistical perspective. 2019-05-30 18:04:44 Overfitting in machine learning due to data set reuse turned out to be less of a problem than feared. There are at least four pieces to the puzzle that explain why. Thread. 2019-05-29 14:18:29 RT @vitalyFM: Turns out overfitting by reusing a benchmark or a test set is a lot harder for multiclass prediction problems than for binary… 2019-05-02 05:13:26 @dr_kuhalla The Burley D'Lite has served us very well. Thule is more expensive, heavier, and has less cargo space. 2019-05-01 00:32:42 @lrntzrsc @jblumenstock Done. Search is closed, folks. This was quick. 2019-05-01 00:19:20 @lrntzrsc @jblumenstock Fairly flexible, but one project should be on privacy in real-world machine learning applications. 2019-04-30 23:58:57 Josh Blumenstock (@jblumenstock) and I have funding for a postdoc at UC Berkeley. Email us for details if you're interested. 2019-04-24 17:56:33 The impact @mer__edith has in AI ethics is second to none. What happens to her at Google will be a gauge for the wellbeing of the entire field. Watch closely. 2019-04-24 01:16:07 I love it when researchers whose work I admire for its depth and rigor make these admissions. It makes technical research feel an awful lot more human. https://t.co/E5wOdU3640 2019-04-19 16:40:26 @emilymbender Kudos for anticipating this and writing a thoughtful note. 2019-04-19 15:24:11 RT @emilymbender: I've seen several different #NLProc folks suggesting today that it would fun/interesting/worthwhile to use BERT or GPT-2… 2019-04-18 15:11:00 Reviewer: Can you be more specific about the methods used to establish your claim? Me: https://t.co/YIPamksJVq 2001-01-01 01:01:01