2022-03-12 08:11:00 CAFIAC FIX 2022-01-19 20:32:19 @vborghesani @fpedregosa @unige_en @NCCR_Language Congratulations Valentina! 2022-01-18 00:19:16 RT @nhannahjones: I was invited to give an MLK speech today and a small number of members of the group hosting me wrote and then leaked ema… 2022-01-17 08:11:00 CAFIAC FIX 2022-01-12 21:42:30 We now aim to expand this to actor-critic techniques, where the loss for the critic relates to the performance of the actor, and where we have convergence guarantees regardless of the parametrization of either the actor or the critic. 2022-01-12 21:42:29 More importantly, we get convergence guarantees *in the general function approximation case*, regardless of the parameterization chosen. This is true for *any* number of inner loop optimization steps and also gives guidance on the regularization parameter. https://t.co/Ha3ysNVmDF 2022-01-12 21:42:28 A particular choice leads to a variant of PPO that does not use importance weights and has the KL in the "easy" direction. This leads to consistently more stable performance on MuJoCo. https://t.co/oqCAXTkXUJ 2022-01-12 21:42:27 It turns out that these operators stem from performing optimization, more specifically mirror ascent, in a functional space. By varying the mirror map and the functional space, we generate a whole family of policy gradient algorithms, recovering popular ones as special cases. 2022-01-12 21:42:26 I am pleased to announce that our work on designing general surrogate functions for policy gradient methods with convergence guarantees is now out on arXiv: https://t.co/KpDuLOtRKa . This is work done by @sharan_vaswani and many other great folks. 2022-01-11 08:11:00 CAFIAC FIX 2022-01-04 16:24:14 So it was @timnitGebru 's courage that empowered me and I am hopeful that the likes of her and @willie_agnew will encourage established researchers to get out of their comfort zone and stand up for what they believe in. 2022-01-04 16:24:13 One good thing to come out of this shitshow is that it once again highlights courageous young researchers who, despite their precarious position, stand for what they believe in. Congratulations and thank you, @willie_agnew . https://t.co/gfXgzaQXZx 2021-12-30 22:36:52 RT @Abebab: Hard to imagine but it's possible to be optimistic about AI & 2021-12-27 08:20:00 CAFIAC FIX 2021-12-20 21:08:43 It is not too late to apply (but better sooner than later)! If you have questions about any opening, reach out to me (here or at nicolas@le-roux.name). https://t.co/AFPuDx0Ues 2021-12-18 22:46:09 @pcastr @Twitter Do you use Adguard? I heard it could be the culprit. 2021-12-16 15:28:18 RT @o_saja: Bumping this up ICYMI! @MSFTResearch Montreal is hiring at different career levels (postdoc + senior or principal researcher)… 2021-12-12 19:22:58 RT @hugo_larochelle: Today, @RaiaHadsell, @kchonyc and I are happy to announce the creation of a new journal: Transaction on Machine Learni… 2021-12-10 15:09:12 RT @jennwvaughan: Happening today! I'm happy to chat about research or whatever is on your mind. Table 8. My @MSFTResearch colleagues @Joh… 2021-12-09 14:36:19 @Abebab Maybe you underestimate the impact you and a few others had on people like me. I feel like many seeds were planted and it will take some time to sprout but this is still cause for hope. Happy to chat more with you about this. 2021-12-09 14:32:55 @Abebab There is a growing group of talented people who critically look at the promises made by AI proponents and teach all of us how to contextualize and to assess the societal impact of our work. This group becomes more and more mainstream and that gives me hope. 2021-12-09 14:28:11 @david_madras Hi David, we are hiring at various levels at MSR Montreal: https://t.co/AFPuDx0Ues 2021-12-06 15:58:49 RT @o_saja: Very last call for FATE MTL @MSFTResearch 2022 internships! @NeurIPSConf folks if interested in issues re: fairness/transparenc… 2021-12-03 21:14:48 And because everyone's career is different, we do not require a PhD for these positions as long as you can prove your research experience. More generally, here are other positions with our friends and colleagues in NY which do not require a PhD either: https://t.co/st2WM0zSTc . https://t.co/QM2HMe84g5 2021-12-03 20:37:14 - https://t.co/yQeyKoeMwQ , where we decompose the reward function in several heads, which we train independently. This simple trick allows us to reach a performance at Ms Pacman that is undefeated at this date. 2021-12-03 20:37:13 3/ Reinforcement learning - https://t.co/g8iDe5zquT , where we develop a practical algorithm for offline RL with safe policy improvement guarantees in finite MDPs. 2021-12-03 20:37:12 - At MSR, you can also work with product teams and https://t.co/gaHpkwnO9I, published at CHI, is a good example of the more practical work we do to map possible issues in deployed systems. 2021-12-03 20:37:11 This was done in collaboration with @katecrawford (author of "Atlas of AI") from MSR New-York. 2021-12-03 20:37:10 2/ FATE - https://t.co/sZSsWFsSeH is an audit of several benchmarks for fair ML and assesses what exactly is being measured. This work is representative of the lens through which we look at problems with data and deployed systems. With collaborators from product teams. 2021-12-03 20:37:09 In the meantime, I would like to share some representative works from our lab: 1/ DL and language - https://t.co/1ZWDQbSQzc represents the span of our interests in language: interaction, grounding in an embodied/visual world, how language can facilitate generalization. 2021-12-03 20:37:08 There are different links to apply for Senior and Principal but feel free to apply to both if you do not know which one applies to you. The team's current research interests span FATE, deep learning, NLP, and RL. We are seeking someone whose research agenda would complement ours. 2021-12-03 20:37:07 .@MSFTResearch Montreal is hiring a permanent researcher at either the Senior or Principal level: Principal: https://t.co/fvLaVt6OOp Senior: https://t.co/VwwfUxXfIa 2021-12-03 14:18:45 @criticalneuro @MSFTResearch @hannawallach Congratulations Ida! 2021-12-02 14:11:33 RT @nitashatiku: I talked to @timnitGebru about her new AI research institute, launched on the one year anniversary of being fired from Goo… 2021-12-01 22:02:17 @SebastienBubeck @geoishard Congratulations Seb. 2021-12-01 21:15:37 I got to interact with @jennwvaughan as an SAC during this year's review period. The level of care and thoughtfulness she displayed to make sure the fairest outcome was reached was astounding. I am in awe of the work she (and I'm sure the rest of the PC) put in. https://t.co/HI54LnbGac 2021-12-01 01:26:48 RT @marcgbellemare: Congratulations to @agarwl_ and @max_a_schwarzer for their Outstanding Paper Award at #NeurIPS2021! It was a thrill to… 2021-12-01 00:44:52 @pcastr @agarwl_ @max_a_schwarzer @AaronCourville @marcgbellemare Congratulations, what a remarkable team. I look forward to reading your paper. 2021-11-30 15:19:36 RT @AJLUnited: Have you supported #AJL this #GivingTuesday? Be sure to post everywhere about it with our "I Support AJL, I Support Algori… 2021-11-16 19:46:32 @alexhanna @JessieNYC I mean, writing such a personal email must have taken Lynda a lot of carriage. 2021-11-12 19:46:41 @soldni @QueerinAI Done, thank you for your work. 2021-11-10 19:34:01 @BachFrancis Average researcher: gets a 2-minute introduction by the event organizer. Francis: gets a 30-minute introduction by the French president. 2021-11-08 21:51:14 I had a lot of fun collaborating with @sharan_vaswani , whom you might know from such hits as https://t.co/YaijyV3XCE and https://t.co/iiIX4GlJbV . If you are a student interested in combining theory and practice and want to discover beautiful BC, this is a great opportunity. https://t.co/sxr4bFaUVm 2021-11-06 23:20:00 CAFIAC FIX 2021-11-06 19:50:00 CAFIAC FIX 2021-11-06 18:59:00 CAFIAC FIX 2021-11-01 19:20:00 CAFIAC FIX 2021-10-07 03:16:59 @ST4Good @MSFTResearch @susan_dumais https://t.co/WO7w0nkKgw 2021-10-05 19:04:49 @yaroslavvb . @fpedregosa and Courtney Paquette have been working on this for the last few years and got some really nice results, for instance https://t.co/M55sb9NltH and https://t.co/q3akWaRCe9 . 2021-10-05 16:26:02 I am ecstatic to start a new chapter at @MSFTResearch in Montreal where I'll get to work with many great colleagues under the leadership of @susan_dumais. I plan to work on model reusability and decentralization of power and we'll soon post about available positions. 2021-06-09 19:49:48 @MarlosCMachado @AmiiThinks @CIFAR_News Congratulations Marlos! 2021-06-09 19:49:48 @MarlosCMachado @AmiiThinks @CIFAR_News Congratulations Marlos! 2021-06-08 14:02:43 @RaquelUrtasun @Waabi_ai Congratulations Raquel! 2021-06-08 03:04:35 @shibl Thank you Shibl. I wish you the best! 2021-06-08 03:03:40 @thomas__coffin Thanks Thomas, although let's be honest: this decision is much easier for me than it is for those in more precarious positions. I have a visa, no financial problem and Google did not smear me. We need people like you to protect those who do not have the same privileges as me. 2021-06-07 01:13:13 I also recommend Race After Technology from @ruha9 (and pretty much all of her work), as well a wonderful article from Green and Viljoen (https://t.co/Qf2eX4lKnY). 2021-06-07 01:13:12 I highly recommend Atlas of AI, especially to those who wonder what their responsibility as engineer/researcher is. It definitely features prominently in the list of works that helped me get a better understanding of my social responsibility. https://t.co/v54eJh4cU3 2021-06-04 15:44:32 Notice the importance of accountability, echoing @timnitGebru 's email from last year (from https://t.co/arUlom4kqf). When people say they care about X without saying how they will be held accountable, they mean "I care unless it costs me something". https://t.co/5Pe8EhOiG9 https://t.co/yKrAA9mU69 2021-06-02 20:01:44 @SashaMTL Merci Sasha! 2021-06-02 20:01:29 @alexhanna @timnitGebru I'm glad one of the last things I could do at Google was help them see your thoughtfulness and dedication to do the right thing. 2021-06-02 20:00:00 @cazencott Thank you! 2021-06-02 19:59:41 @tejuafonja Thank you Tẹjúmádé. I look forward to seeing all your future accomplishments. 2021-06-02 19:59:02 @aharutyu Thank you Anna. 2021-06-02 19:58:34 @JessicaSchrouff Thank you for the kind words, Jessica. 2021-06-02 19:58:06 @panfordkobby I remember! How did it go? 2021-06-02 19:56:57 @timnitGebru Thank you, my friend. One of the main drivers behind my actions was witnessing your courage and your relentless pursuit of justice. You paved the way, all I had to do was walk it. 2021-06-02 19:55:20 @sguada Sergio, I cannot overstate how many things you taught me. I hope you'll have more success than I did changing the culture from the inside. 2021-06-02 19:54:17 @celestekidd Thank you Celeste. It means a lot, especially coming from you. 2021-06-02 05:28:01 @MarlosCMachado Thanks Marlos. It was a great pleasure working with you and, along with @its_dibya , you are a co-author of my favourite uncited paper :D 2021-06-01 23:52:41 @hugo_larochelle Thanks Hugo. You assembled a team of talented, thoughtful and caring researchers and engineers in Montreal and it was a delight to get to work with them every day. 2021-06-01 20:43:28 @pcastr I am privileged and humbled to have been your colleague, Pablo. You've been an inspiration and I hope our paths will cross again soon. 2021-06-01 19:38:26 I left Google last Friday. I will sorely miss my incredible colleagues and hope they one day get to enjoy an environment where they feel safe to take risks, be themselves and challenge the status quo. I have not decided what I will do next. I will take a break and rest for now. 2021-06-01 12:26:03 RT @kchonyc: i've received the Ho-Am Prize (Engineering) which comes with a pretty substantial cash prize. i plan to spend it for a few cau… 2021-05-29 18:31:17 RT @mer__edith: This. Including the colleagues who suddenly stop replying. The people whose work engages "justice" who politely explain tha… 2021-05-28 12:40:13 From my (albeit limited) experience, the price you end up paying if you speak about this from a position of power is not your safety nor your job but discomfort. And the more you do it, the lower that price, so it's best to start now. From an email I sent to Brain some time ago: https://t.co/w8PQUUxWXq https://t.co/LgeJ0HEs8q 2021-05-27 19:02:38 The art teacher asked us to bring a box of 30 colored pencils. I must have lost one because my box had only 29. I decided not to bring the box at all since the requirement wasn't met anyway. I was 11. https://t.co/znC5cciAuf 2021-05-27 18:16:29 .@dtarlow2 introduced me to this great article on the topic, "Culture is the Behavior You Reward and Punish": https://t.co/mFVGCPbKbV . "When all the “successful” people behave in the same way, culture is made. " https://t.co/lDr5TnHeMX 2021-05-27 12:02:40 RT @StarFeuri: Our paper "Oh No, Not Another Trolley!" has been accepted to @IEEESSIT. We survey CS majors about their exposure to ethics i… 2021-05-26 21:28:49 @BeEngelhardt @Princeton Congratulations Barbara! 2021-05-25 15:13:47 @mark_riedl I shared with them how much vacation days I had (almost 7 weeks between paid time off and holidays) and I told them to take at least as much. 2021-05-20 13:41:12 @venikunche So far, I always refused. This has not prevented me from getting offers but I realized I am in a privileged position. 2021-05-17 21:06:45 RT @AJLUnited: Here are the "Face Queens" — @jovialjoy, @timnitGebru, and @rajiinio that @60Minutes erased during their episode on facial r… 2021-05-17 19:49:39 RT @timnitGebru: So strange to see this from @60Minutes. I was watching the episode and wondering how they didn't have @jovialjoy at least,… 2021-05-14 03:54:04 @doomie To explain pi, you often have to circle back, whereas Cantor's diagonal is more about continuous effort. 2021-05-14 03:46:24 @doomie For the latest one, this is a fun entry point to discuss the multiple infinites. Would Cantor's diagonal argument be accessible to a 5yo? 2021-05-13 16:27:03 @NicLoizou @JohnsHopkins @HopkinsEngineer @JHUCompSci Congratulations Nicolas! They're lucky to have you. 2021-05-13 03:04:05 @chrisalbon But your book is a principal component of that list. 2021-05-10 18:37:45 My question around scholarships was so poorly worded that I deleted my tweet to avoid further confusion. I take good note of all the replies, though. 2021-05-10 15:57:11 @ReyhaneAskari This is a completely hypothetical case and I'm asking for a friend anyway. 2021-05-10 14:23:40 @amy_tabb I pay the tuition fees and conference costs in addition to the full stipend. 2021-05-10 14:19:35 @jasonaholliday The stipend is reduced by X/2$ for the duration of the scholarships so the "effective" stipend of the student is only increase by X/2$, instead of the full X$. Pinging @bouzoukipunks to make sure everything I say is an accurate description 2021-05-10 14:18:14 @digpath @bouzoukipunks Entirely my fault for not being clear enough. I guess it didn't cross my mind that that could be a possibility. 2021-05-10 14:15:42 Retweeting this important clarification from Ioannis: Keeping half of the scholarship allows me to have more students but I fail to appreciate how big of an impact it can have on the student. https://t.co/sajTnN3cyg 2021-05-10 14:12:44 @bouzoukipunks Yes, thank you for the clarification, Ioannis. The brevity of a tweet made me omit important information. 2021-05-10 14:11:58 This is an example of such a scholarship: https://t.co/Y9Ki2vQQ0J . 2021-05-10 14:10:57 @amy_tabb Thanks, I changed the wording from "grant" to "scholarship" as I think it's a better description. 2021-05-10 14:09:19 Common practice where I am a professor is that, if a student is awarded a scholarship worth X$, they get X/2$ and the advisor gets the other half. Do you think the student should get the full amount? Standard PhD stipend is 54% of local median full-time income. 2021-05-07 20:21:06 @doomie OK, I'll schedule a meeting with you. 2021-05-05 19:46:19 @doomie I showed this paper to Arthur Szlam in 2010 and they told me it was a trivial, well-known result, reproducing it in 3 lines of Matlab. IIRC, this was a direct application of harmonic analysis. 2021-05-05 11:46:29 @rishiyer @NeurIPSConf This does not replace a deadline extension but they can reach out to one of the many people who offered their help for this deadline: https://t.co/9SeJwZAwGK . 2021-05-03 19:26:38 Bumping this for visibility. Anyone can be a mentee, don't hesitate to reach out to anyone on the list. Potential mentors: there are far fewer of you than last year (I didn't automatically include anyone) so don't be shy! https://t.co/vJOZPacTPK 2021-05-03 18:15:06 @neu_rips @AlexShtf But that would be going on a tangent and you're above that. 2021-05-01 21:56:51 RT @abbyfuller: https://t.co/Zpd8Z1IlgN 2021-05-01 18:28:03 Please feel free to reach out to me if you used it as it will help me gauge how to adapt it for the future. 2021-05-01 18:26:58 I took into account some of the feedback from last year. In particular, if you keep the link to edit your answers, you can now remove your name at any time. I always welcome feedback if you feel this could be improved. 2021-05-01 18:23:35 The response last year was overwhelming so we're doing it again this year, but with more planning: NeurIPS 2021 submission mentorship! Since we had fewer requests than mentors last year, this is open to everyone who feels they would benefit from it. 2021-04-14 18:03:49 RT @fpedregosa: New blog post: Acceleration without Momentum. After two blog posts on momentum, now one on how to get the same effect witho… 2021-04-11 01:37:40 RT @sindero: @mmitchell_ai Umm. Right here. Significant contribs to DL since BEFORE it was DL. Notable R& https:/… 2021-04-06 19:12:25 I know Samy wanted to do everything in his power to make the environment more welcoming and safe to those who wanted to stay / could not afford to leave. That someone who fought so much decided to call it quits is telling. https://t.co/9tIM9Lltiu 2021-04-05 18:54:19 RT @BachFrancis: It is not about chocolate, but I hope you will find it comforting. I am writing a book about learning theory. See more det… 2021-04-01 20:10:49 @frank_e_curtis This is such a nice paper. I keep recommending it to everyone who wishes to have an accessible yet thorough understanding of the field. Congratulations! 2021-03-25 12:12:08 RT @geoffnegiar: Hi #ml #optimization twitter! Recently @fpedregosa and I have been working on a *constrained* optimization library for Py… 2021-03-24 17:26:44 @gautamcgoel If it doesn't, can't you just merge the entries in Google Scholar? I've done this many times for multiple versions of similar works. 2021-03-21 01:32:29 RT @dinoman_j: https://t.co/CXcMNeq0pl 2021-03-18 14:39:50 @mpd37 It probably doesn't matter much but I think you do great research and I often go back to your survey on policy search. 2021-03-17 12:45:17 RT @rajiinio: There's something important that no one seems to be saying about this article so I'm putting my thoughts here to get it off m… 2021-03-16 20:09:09 RT @Abebab: Last Wednesday, on the 10th of March-2021, @vinayprabhu & 2021-03-13 23:41:09 @vborghesani @fpedregosa @BouhaliFlorence We did the same thing today! Save for the run. 2021-03-13 15:21:24 RT @glichfield: With all the pushback from Facebook against @_KarenHao's recent story, and as one of the editors on the piece, I thought it… 2021-03-12 21:55:56 @lawrennd @pcastr @_KarenHao I'd be interested in having a better understanding of the different levers to shift the balance of power. Do you have pointers to a good introduction? 2021-03-12 17:46:45 @lawrennd @pcastr @_KarenHao How could I disagree with you, Neil? 2021-03-12 17:46:21 @lawrennd @pcastr @_KarenHao But this requires legislating and I feel this will be an easier discussion (and stronger legislation) if the people with power have first-hand experience of the negative impacts of disproportionate corporate power. 2021-03-12 17:40:42 @lawrennd @pcastr @_KarenHao I agree with you if you meant "This is not enough", just like I think we agree that the solution requires decentralizing power (whether it means data, capital, or sth else). 2021-03-12 15:22:53 @pcastr @lawrennd @_KarenHao True, and I want to be clear that "we" refers to the loudest voices. There has been a movement to diversify VC funding for a long time, it just does not have nearly as much visibility as the current narrative (or is cast as "identity politics"). 2021-03-12 15:14:25 @lawrennd @_KarenHao Could it be that all these companies are run by people who have little or no first-hand experience of the struggles that this obsession over growth amplifies? If so, this is not so much a characteristic of large companies than of the founders we (as a society) decide to fund. 2021-03-11 16:23:00 .@_KarenHao and @kharijohnson are my favorite reporters when it comes to the broader impacts of AI. This is a great piece from @_KarenHao, showing once again that our impact does not necessarily align with our intentions, however good they may be. https://t.co/3qzg9kJ4BB 2021-03-08 21:44:46 RT @michaelharriot: I have received a lot of emails pushing back on this quote from my mother. “A black child’s humanity can never be ful… 2021-03-08 00:50:49 @timnitGebru Replies I got when raising google issues: - "You see everything with a negative lens." - "Have you tried not caring about it?" 2021-03-05 17:15:43 RT @xkianteb: What is the "right" embedding space for prediction, reinforcement learning, imitation learning, and planning? We try to tackl… 2021-03-04 17:25:10 @mariadearteaga @sinafazelpour It was a great tutorial, thank you. 2021-03-04 13:13:59 @Abebab "The production of audit data about a social problem must be accompanied by narrative tools that help to drive change: “[d]ata, in short, do not speak for themselves and don’t always change hearts and minds or policy”", @red_abebe et al. citing @ruha9 in https://t.co/G8RxzakpyN 2021-03-01 22:03:03 @mekkaokereke I forgot to add a source: https://t.co/Zx0PEQpGaN (just learnt it had been replaced with another tax in 2017). 2021-03-01 17:32:34 @mekkaokereke That's in place in France, by the way (starting at about $2 million). 2021-02-25 17:06:07 @RaquelUrtasun @Uber @dkhos Congratulations on what you accomplished there. I'm excited to see what you will do next (with no doubt that it will be impactful). 2021-02-19 22:11:55 @mmitchell_ai Wow. Even though I'm sure you'll rise even higher, this is terrible news for you and your team. I am so sorry. 2021-02-17 14:19:54 RT @bremen79: New blog post: Perceptron is Not SGD: A Better Interpretation Through Pseudogradients Pseudogradients (Polyak& 2021-02-11 20:47:51 RT @katecrawford: I'm so excited that my new book is out in April: ATLAS OF AI: Power, Politics, and the Planetary Costs of Artificial In… 2021-02-11 04:07:39 I've clearly worked less since I had a kid and, whatever impact this had on my career, I don't regret it one bit. 2021-02-11 01:21:28 You don't have to do long hours to do meaningful work. Research can be just a job if that's what you want. Your self-worth is not tied to your professional success. It is possible to have a bigger impact enabling others than doing your own research. https://t.co/S7kyxNDEqw 2021-02-09 15:14:12 RT @dumoulinv: I'm helping organize an ICLR workshop on learning in the context of limited and adverse data. If your work fits this topic,… 2021-02-01 21:44:32 @kareem_carr Mike Jordan had a great lecture on that question: https://t.co/47TqCfvVn3 . 2021-01-30 21:44:31 @kchonyc @an_open_mind This is what we did for ICML. We wanted to promote different aspects, choosing one or several papers for each. 2021-01-27 21:41:28 @doomie 1/ As someone who wants to limit their flying, I hope we'll have the possibility of a great remote experience. 2/ I'm less sure about Australia, but given NZ's stance, highly doubt that it'll happen. 3/ There might be more suited words to use than "crazy" 2021-01-25 16:28:09 RT @WillOremus: "In more than 60% of cases documented by researchers, GPT-3 created sentences associating Muslims with shooting, bombs, mur… 2021-01-25 15:02:26 Pablo is so talented, it's a blessing to enjoy all that he's creating. https://t.co/vH9Hk6omSH 2021-01-24 15:27:23 @beenwrekt @neu_rips @SoloGen I think Damien Scieur has results where IA really helps. 2021-01-24 15:20:39 @neu_rips @SoloGen @ylecun @sharan_vaswani But I moved away from optimization specifically because I wasn't excited about this pursuit of rates anymore. I find other questions such as how to do stochastic line search or how to use stale gradients much more interesting because they make optim faster/easier in practice. 2021-01-24 15:18:04 @neu_rips @SoloGen @ylecun @sharan_vaswani Two cases where I find them relevant: 1/ Interpolation, where you can show it's almost as if nonconvexity didn't matter, enabling things like line-search. 2/ When they are used in inner loops, like solving a prox, where accurate solutions are critical. 2021-01-24 15:15:43 @neu_rips @SoloGen @ylecun @sharan_vaswani I'm a big fan of optimization theory in the convex case because it is a really good predictor of empirical performance. I am much less confident about the relevance of rates in the nonconvex regime, save for some specific cases, which I think aligns with your original sentiment. 2021-01-24 15:13:19 @neu_rips @SoloGen @ylecun @sharan_vaswani Given that our understanding of the dynamics of stochastic (and even batch) methods in the nonconvex regime came long after practitioners used them successfully, it might be they implicitly accepted that rates don't tell the full story, which seems fair. 2021-01-24 14:46:05 @neu_rips @SoloGen @ylecun Even better, when the network is so big that you reach interpolation, constant stepsize without averaging gives you convergence to the solution, at a linear rate if you satisfy the PL condition, see Theorems 3 and 4 of https://t.co/MqCZMWfZ9t by @sharan_vaswani et al. 2021-01-24 14:44:25 @neu_rips @SoloGen @ylecun On the other hand, non-iterate averaging gives you convergence to a neighborhood of a stationary point, for instance Theorem 4.8 of https://t.co/enreWNNzKO . 2021-01-24 14:43:48 @neu_rips @SoloGen It's a bit more subtle than that. IIRC, @ylecun was using it a long time ago. The issue with iterate averaging is that the non-asymptotic rate can be slow and you carefully need to tune when to start averaging. 2021-01-23 22:16:23 @black_in_ai @DeepIndaba .@ermgrant and @le_roux_nicolas* 2021-01-22 13:50:53 "Nothing about us without us". Indaba offers rightful opportunities to the African ML community and is governed by African ML researchers. Consider signing up as a mentor or mentee. https://t.co/PUXbdcXYpB 2021-01-21 22:05:38 I look forward to reading it. https://t.co/vwaq8etCBF 2021-01-21 20:33:44 RT @mmitchell_ai: Hello, world! Similar to @timnitGebru, I abruptly lost access to my professional calendar. Any upcoming talks, deadlines,… 2021-01-21 20:33:17 RT @_KarenHao: When a company like Google or Microsoft or OpenAI builds these gargantuan models so that *they,* our benevolent tech dictato… 2021-01-20 02:36:18 @databoydg @timnitGebru I remember she raised her voice at a meeting 18 months ago. She had it coming. 2021-01-20 02:31:12 RT @timnitGebru: She's not fired **yet**. But apparently they've told her she will be locked out for at least a few days. https://t.co/H1ib… 2021-01-19 17:45:19 @fulhack @ogrisel .@ogrisel is a star and a wonderful person to have in the community. 2021-01-19 16:59:24 RT @franklinleonard: “Check yourself before you wreck yourself.” - Martin Luther King Jr. https://t.co/R5aQ3Ernrx 2021-01-17 22:54:40 @thomas__coffin Never attribute to maturity what can be explained by laziness. 2021-01-17 22:35:58 To my dismay, this is not a brachistochrone curve. 2021-01-17 22:32:32 This was a productive weekend. https://t.co/7PwsmXc9sp 2021-01-16 03:24:00 @MaxKaminCross @1nf0_s3c @jackclarkSF @signalapp And to @mer__edith who amplified your call 2021-01-16 02:37:24 @MaxKaminCross @mer__edith @signalapp Well, you said you matched up to 1k USD. What was I supposed to do? 2021-01-16 02:34:56 @MaxKaminCross @mer__edith @signalapp Nor to conjugate, apparently. 2021-01-16 02:34:29 @MaxKaminCross @mer__edith @signalapp Ahah, I'm not even able to type my name correctly. 2021-01-16 02:34:00 @MaxKaminCross @mer__edith @signalapp I didn't know what the exchange rate is so decided to err on the safe side. https://t.co/nk9463iZii 2021-01-15 23:33:30 @moskitoes_bite @NandoDF I'm thinking of workshops which are great places to meet mentors and collaborators. For instance: - https://t.co/9NaJnIEjHi - https://t.co/GRYspZkNOy - https://t.co/qnfcHUc9Ds 2021-01-15 21:45:45 @cottascience @NandoDF @black_in_ai A more concrete example: @ermgrant and I hastily set up a mentor list for @NeurIPSConf last year (https://t.co/tuqkqHvyA4). However, this list does nothing to address the deeper challenges faced by many. This is why we are working w/ affinity groups to hand the list over to them. 2021-01-15 21:44:05 @cottascience @NandoDF Agreed. Let me rephrase: these are all actions taken by, e.g., @black_in_ai and their relevance was not obvious to me before talking to them. Hence my deference to their expertise to know what's best. 2021-01-15 21:42:16 @moskitoes_bite @NandoDF If you feel no need to be represented by WiML but would rather find more theory-minded researchers, it seems like channels already exist for that. Am I mistaken? 2021-01-15 21:41:27 @moskitoes_bite @NandoDF Agreed but @NandoDF 's tweet was specifically about increasing representativity. WiML supplements rather than replaces general venues in order to address the structural issues limiting the number of women in the field. 2021-01-15 21:34:15 @NandoDF For instance, since you mention students in poor countries, barriers can be: - helping w/ application fees to go into grad. programs - training for GRE or SAT - actual research mentorship But I would be hard pressed to tell you which of these actions has the largest potential. 2021-01-15 21:24:32 @NandoDF This can mean financing, taking on non-leadership roles or advocating for them to your company or institution. 2021-01-15 21:15:15 @NandoDF Discussing with the affinity groups made it abundantly clear that they have a far better understanding of what works, what the needs are (vastly different from one person to another) and how to best act than I will. So, to me, what works best is to support them. 2021-01-15 17:31:28 @balazskegl @Abebab Great questions. Let me draft a reply that I will then post in a twitter flood as I usually do. 2021-01-15 16:11:31 Over the last few years, I got less interested in pure technical achievements and more in the societal impact we wanted to have. This remarkable piece by @Abebab clearly explains why we need to move away from tech-solutionism: https://t.co/05Oabd0Avt . 2021-01-15 01:12:05 RT @katecrawford: I've known Laura Poitras for years and her integrity and brilliance are unparalleled. Without her, First Look wouldn't ex… 2021-01-14 14:51:10 @leocasarsa @_LXAI - Marlos C. Machado: @MarlosCMachado - Cristina Vasconcelos: @Cristin13377923 2021-01-14 14:12:04 RT @dinoman_j: https://t.co/Hu1oZtWEXc 2021-01-05 15:39:37 @pcastr It sounds like a Disney song. In case you don't know about this album: https://t.co/4mHq2rk96B 2020-12-23 20:50:24 RT @celestekidd: Moving AI Forward event starts in 15 minutes, and is free to join. Looking forward to seeing many of you there. https://t.… 2020-12-22 17:05:26 @alexhanna Unity is great but freedom is better: https://t.co/qZEtQXyr3o 2020-12-21 23:28:27 RT @RealAbril: I'm finna tell yall why @Google fired me- their MOST successful diversity recruiter in the history of their company- with t… 2020-12-21 22:38:50 RT @fpedregosa: Here comes the second part on my series of blog posts on the connection between optimization methods and polynomials. Now i… 2020-12-21 19:04:20 RT @math_rachel: People who consider themselves: against racism, but hold Black women to different standards/processes against sexism, b… 2020-12-21 19:04:13 RT @timnitGebru: This is what always happens. Everyone says they’re against these things but when met with the specifics they either “both… 2020-12-21 14:52:04 RT @bneyshabur: If you believe you are at a disadvantage in ML community (e.g. because of your race, gender, nationality, background or oth… 2020-12-18 03:56:41 RT @erikashimizu: PSA: https://t.co/4zsSAsIHuj 2020-12-17 20:50:18 @ylecun @mlittmancs @BrownHCRI @isbellHFh I was baffled by @isbellHFh 's acting ability and his mastery of the "pondering gaze". 2020-12-17 03:15:51 This should have been "unconscious bias". Our current work with @FartashFg and @fpedregosa gave me a brain fart. 2020-12-17 01:09:08 @wicsthrowaway The tension of the BLM movement is the direct consequence of the inaction following Colin Kaepernick's peaceful protest. However uncomfortable it is to witness this tension, I can attest it is nowhere near the pain that caused it to erupt. 2020-12-17 01:03:29 @wicsthrowaway Although we'd all wish for change to happen smoothly, history tends to show that it is rarely the case. Tension arises out of necessity and it falls upon all of us to create the conditions where this tension would be unnecessary. Here is MLK's take on tension: https://t.co/gzqXY0BPpH 2020-12-16 19:41:35 @shibl Ah, interesting question. But "motivated" indicates an intention. While I appreciate why the intention matters, I am more interested in the consequences. Ultimately, do these events amplify the marginalization of certain groups? Based on the testimonies, I am convinced they do. 2020-12-16 19:02:19 I am now careful with my reviews on pubapprove. https://t.co/2HS9B8S2ex 2020-12-16 18:55:34 @shibl But you're right that this plays for him as well. He's thoughtful and surely he must have foreseen the consequences so this must have been really bad. Despite this, I'm still way more convinced by Timnit's account, mainly because of the sheer number of converging testimonies. 2020-12-16 18:52:00 @shibl Which responsibility exactly are you referring to? From what I saw, Jeff still refuses to acknowledge he fired her. 2020-12-16 18:50:46 @shibl Here, I believe the potential impact for marginalized groups in tech weighted by the probability that Timnit says the truth far outweighs the impact for Google weighted by the probability that Google tells the truth. 2020-12-16 18:49:25 @shibl The opportunity of whether or not to speak w/o all the facts depends on the consequences. I'd never take such a stance at a criminal trial. The level of certainty needs to be commensurate with the sentence. 2020-12-16 18:47:24 @shibl True, but we'll never know them. So what do we do? We can stay silent, but this is effectively siding w/ Google here. Or we can take a risk based on our estimated posteriors and endure the discomfort of knowing that we might be wrong. 2020-12-16 18:04:54 Here are the resources if you want to know more about Timnit's side of the story: Her interview: https://t.co/WvGKNK5bBg The timeline of events: https://t.co/AnDr79iQZp Thread by @shaft on Google tactics: https://t.co/wy2dzWiP7T 2020-12-16 17:31:08 I have more respect for someone saying "I took the time to actively engage and listen to both sides, and I ultimately concluded that Google did the right thing" than for someone saying "I do not know Timnit but Jeff is a great guy so I tend to believe him." 2020-12-16 17:31:07 This is what "Assume no intentions, observe all impact" means. To enable racism and sexism, you don't have to believe in an inherent inferiority of a race or gender. You don't even need to have implicit bias. You just have to be complacent and never question those you love. 2020-12-16 17:31:06 Jeff is someone I look up to. He's achieved more than I ever will and always was a great colleague and mentor. He's also been genuinely involved in efforts to diversify ML. From what I know about him, I'd never believe that he would do something as damaging as firing Timnit. 2020-12-16 17:31:05 Someone I look up to told me that they remained silent about recent events because they knew very little about what happened, only that Jeff was a great guy who had been an advocate before. But they didn't want to be a white guy supporting another white guy so they stayed mum. 2020-12-16 14:31:38 RT @_KarenHao: After a long two weeks since being forced out by Google, I sat down with @timnitGebru to reflect on her time at the company,… 2020-12-15 19:34:16 @dlowd @pmddomingos @AnimaAnandkumar Would these have been Stalingrades? 2020-12-15 03:53:27 @timnitGebru @dlowd Can I join? "Twitter mob", "virtue signalling", "this reminds me of the soviet union", ... Gotta catch'em all! 2020-12-13 19:24:57 @vborghesani @fpedregosa Is it me or is someone having more fun than the other? 2020-12-11 23:20:26 @TalKachman @SebastienBubeck Convex Optimization: Algorithms and Complexity https://t.co/8uAYvt88yO 2020-12-11 21:53:11 There I am, having fun with mirror descent and policy gradient, and I cannot recommend @SebastienBubeck 's introduction to mirror descent enough. His book is super clear and gives you the impression you know everything you have to in under 15 (small) pages. 2020-12-11 17:08:25 @math_rachel @ResistanceAI Some NeurIPS attendees, who can follow the livestream, are leaving the Zoom meeting to make space for others. Maybe try again? 2020-12-11 15:47:11 .@renireni talks about how white feminism struggles with the concept of whiteness. And nearly all French people I talked to told me that discrimination was not based on gender nor race but on wealth, as if it the three didn't overlap and couldn't compound each other. https://t.co/NYluXa412x 2020-12-10 21:05:47 You need to join Zoom for the Q& 2020-12-10 21:02:10 .@white_martha , @ShamKakade6 and I are now having a Q& 2020-12-10 14:18:49 @balazskegl @ChrSzegedy Inspiring. 2020-12-10 00:58:18 .@_LXAI also waives fees for PhD programs. You can donate here: https://t.co/3Lh99UzeND . https://t.co/I4lvNSSiSQ 2020-12-09 22:34:20 .@katecrawford is such an energetic and inspiring person, this was a pleasure to do this with her and with the many participants who kept asking insightful questions. If you couldn't come but still have questions about leadership in academia/industry, please reach out to me. https://t.co/B71a1zUYsc 2020-12-09 22:31:01 I'm not aware of other affinity groups having similar programs but please feel free to correct me and I'll advertise them as well. 2020-12-09 22:31:00 Application fees for graduate programs are a major barrier to the presence of Black people in AI. BAI is working hard to get them waived but many applicants will still have to pay them. Consider donating to BAI to help them pay these fees: https://t.co/cVlB51byKH 2020-12-09 20:26:04 Given that my first memory of David Mackay was him telling me I "obviously don't understand anything about Information Theory", I'm starting to think I'm not the best at first impressions. 2020-12-09 19:40:00 My first memory of Olivier was him and @BachFrancis making fun of me when I told them I didn't need to regularize my SVM. 2020-12-09 19:37:13 Listening to the memorial reminds me of the chance I had to know Olivier while at @CriteoAILab . Not only was he an exceptional researcher, he was generous and I witnessed him doing the right thing in difficult situations without an hesitation. He will be sorely missed. https://t.co/c41W6RFBjV 2020-12-08 16:05:26 That was fantastic, thank you Charles and all those you invited. https://t.co/07c91lbOXJ 2020-12-08 12:52:28 @kchonyc Thank you. It means a lot coming from you. 2020-12-08 12:31:40 @kchonyc I must be dense but I missed what you didn't understand (or disapproved of). 2020-12-07 18:17:36 I'm excited to be part of this tutorial. In addition to the talks and reading list, you can experiment with the concepts introduced by using notebooks written by @_achan96_ , @dhawgupta , Shivam Garg and Abhishek Naik. https://t.co/wrbWzzlzpi 2020-12-07 16:24:15 @beenwrekt @madsjw @HazyResearch Congratulations Ben, well deserved. 2020-12-07 13:59:49 @konstmish @neu_rips Mine had a lot of ingredients. There was also same raw and dissent. 2020-12-07 13:51:27 @mhajabri @iamcardib @black_in_ai @JeffDean Thanks for tagging me, I did it. 2020-12-07 13:50:28 @iamcardib CAD 5150 to Equal Justice Initiative https://t.co/OI29A3kEY3 2020-12-07 13:47:36 @iamcardib Another USD 10 500 to Black in AI https://t.co/Ictd3dQlCt 2020-12-07 13:46:42 @iamcardib USD 10 500 to Black in AI https://t.co/QHlCgsIwAQ 2020-12-06 00:49:39 Samy has been such a powerful force at Brain over the last few years, this is disheartening. https://t.co/5jiqsQZEFJ 2020-12-05 01:10:17 @timnitGebru You can say whatever you want but the reality is that your submission wasn't appr... Oh no, wait, it was approved by a director, nevermind. The approver surely didn't see the scathing internal revi... Oh, "Left some minor comments but LGTM overall." 2020-12-04 21:58:58 @latentjasper @roydanroy Yeah, the subtitles are complete cabbage. 2020-12-04 16:16:24 @roydanroy My birth name is Nicolas, most people call me Nicholas, but really: https://t.co/UOqpKcqf8O 2020-12-04 02:01:49 RT @GoogleWalkout: We stand with @timnitGebru. We call on Google Research to strengthen its commitment to research integrity and to unequiv… 2020-12-03 22:24:19 @julianibarz And I think it's great to use the tool for that too. But then: - It needs to be done for everyone - The guidelines for how this can happen must be clear. For instance, you can enforce that a paper be submitted early enough for internal review. This was never enforced for me. 2020-12-03 22:21:10 RT @TechWithX: google jumped so quick to get rid of @timnitGebru after constantly parading her around as an example of how much they care a… 2020-12-03 21:26:12 @roydanroy From Jeff's comment (https://t.co/arUlom4kqf): https://t.co/aTUKq185ba 2020-12-03 20:54:37 Now might be a good time to remind everyone that the easiest way to discriminate is to make stringent rules, then to decide when and for whom to enforce them. My submissions were always checked for disclosure of sensitive material, never for the quality of the literature review. 2020-12-03 17:29:50 Great initiative by @EmtiyazKhan and others around mentorship. We got a great response last time, I encourage all those who can to register. https://t.co/Il5szFBqQH 2020-12-03 13:41:34 I am incredibly grateful to @timnitGebru for everything she did and for how she made Google and the field a better place despite all the hurdles. Timnit, I admire your courage to always speak up when times are tough and I support you. Google leadership handled this terribly. https://t.co/jDdkcjVdg2 2020-12-02 18:16:11 RT @MarlosCMachado: I'm very excited our Nature paper is out! Back in 2018, in my job talks, I'd say my main goal was to design RL algorith… 2020-12-01 17:19:44 .@pcastr is one of the nicest researchers I know. I'm glad @JS_Obando got the chance to work with him and I encourage others to jump at any opportunity to do the same. https://t.co/uk8azqvL8Y 2020-11-27 12:39:39 @JeffDean That's a huge cabbage collector. I guess that's why I don't see any leek. 2020-11-26 23:33:58 @isbellHFh @kareem_carr Agreed, that's common kurtosis. 2001-01-01 01:01:01