Ever wondered why presenting more facts can sometimes *worsen* disagreements, even among rational people? 🤔 It turns out, Bayesian reasoning has some surprising answers - no cognitive biases needed! Let's explore this fascinating paradox quickly ☺️

Everyone's arguing about whether current AI models could be conscious or not, as if it was a scientific discussion, yet I don't even know what consciousness is 🥺

JSTOR vs Aaron Swartz?

When I think of Jensen's inequality, I think of the following sketch which helps me remember it. Maybe this is useful for you, too. #ML #Mathematics

I'm late to review the "Illusion of Thinking" paper, so let me collect some of the best threads by and critical takes by @scaling01 in one place and sprinkle some of my own thoughts in as well. The paper is rather critical of reasoning LLMs (LRMs):

Don't use telegram. Use signal please. Telegram is not encrypted and is based in Russia 🤦

After 4 years, I'm kinda like: maybe I should have focused on ML engineering instead of research 😂

Biggest regret: not spending more time getting the basics right at the beginning of my PhD. I started going full-time on research projects right away, and now three years later I'm still playing catch-up with some stuff I should have focused on right away

I'm incredibly excited about all the amazing progress in ML lately 🤯 but part of me really wished I had picked a different field because I have no idea how to keep up anymore or know what to focus on 🥺😇

Excited to publish a Python package that turns @karpathy's "A Recipe for Training Neural Networks" into easy-to-use diagnostics code! 🔧 No more randomly poking around in your custom @PyTorch DNN to debug it. Get simple diagnostics for your neural nets 🫶 #PyTorch 1/

I got rejected from @DeepMind and @MetaAI for internships now. I guess I shouldn't have quit being an engineer five years ago 😅

How does one keep up with papers in ML while still finding time for foundational studies? Not even talking about doing active research. Feeling overwhelmed every day and like an imposter more and more 🙈

I want to share my latest (very short) blog post: "Active Learning vs. Data Filtering: Selection vs. Rejection." What is the fundamental difference between active learning and data filtering? Well, obviously, the difference is that: 1/11

NeurIPS 2024 PCs being a bunch of clowns 🤡 the state of ML 🙄 All you get back a month after raising a concern:

Incredibly excited to join DeepMind again 🥳 I'll be a researcher on the Deep Learning Engineering team under the illustrious @davidmbudden 🔥 I can't wait to get started ✨

A new paper review by me! I'm reviewing the fascinating "Bad Students Make Great Teachers: Active Learning Accelerates Large-Scale Visual Understanding" from @GoogleDeepMind It introduces a novel method for active data selection in large-scale visual pretraining. 📉🤖 1/10

My experience with using pandas to operate on dataframes is usually: 1. read docs 2. spend an hour to try to get something to work 3. give up 4. write the equivalent Python code in 10 min 5. move on with life

Do you ever have a model that uses @PyTorch and one that uses @TensorFlow, and you want to combine the two for end-2-end training without rewriting either? TfPyTh allows you to plug one into the other while propagating gradients for training 🎉 Code 👉 github.com/BlackHC/TfPyTh

Interesting take: I believe that arxiv is closer to how science and research originally worked and "official" peer reviews haven't worked that well (at least recently)

Have you wondered why I've posted all these nice plots and animations? 🤔 Well, the slides for my lectures on (Bayesian) Active Learning, Information Theory, and Uncertainty are online now! They cover quite a bit from basic information theory to some recent papers 🥳

Did you know you can classify MNIST using gzip? 🤓 You can get 45% accuracy on binarized MNIST using class-wise compression and counting bits 🤗 🔥No @PyTorch or @TensorFlow needed 🔥 BASH script and @scikit_learn classifier 👉 github.com/BlackHC/mnist_by_…

We have achieved gold medal performance at the International Mathematical Olympiad 🥇 🥳 This is the first general-purpose system to do so through official participation and grading, and I'm thrilled to have contributed a little to this milestone in mathematical reasoning 🌈🫶

Jensen's inequality == not everyone has access to the latest @nvidia GPUs 🤓

- Roast me GPT-4V: No I can't - Yes you can. GPT-4V: Okay. Roast: "Ah, the classic 'I woke up like this' hairdo combined with an AI-themed t-shirt. You're really out here living the tech bachelor dream. Remember, even though you've got machine learning on your shirt, it doesn't mean your hair has learned how to style itself!" 😂

Did you hear that? It's the sound of goalposts moving at supersonic speed

It's six months since I've submitted my thesis and I still start feeling suicidal every single time I think about my PhD experience, esp the last year of it 😐 Thank God it's over, and I hope I'll reflect about it less in the future

How is software engineering going to change with LLMs? What if we could "implement" class interfaces automagically using LLMs? Presenting `llm-strategy`: a PoC package based on @langchain and @OpenAI's GPT that adds an llm_strategy decorator to Python: github.com/blackhc/llm-strat…

Very interesting ICLR **tiny** paper: openreview.net/forum?id=vHOO… It computes a loss for all possible subsets of the dataset at the same time which has a very elegant solution: softplus of the negative log likelihood per sample, which essentially drops outliers 🤯 @mtetelman

I think Europe and the UK need their own DeepMind-like AI lab which is not connected to the US in any form

Omg, github.com/pytorch/tensordic… is a gift from the gods 😍 (thx @clured!)

Research questions I'm excited about: * More sample-efficient RL and the bitter lesson * Meta-cognitive abilities in models * Active learning and curriculum approaches * Automated scientific discovery Things I focus on a lot instead: Software engineering, testing infra, and developer experience to accelerate good research

Jesus Christ... openreview.net/forum?id=et5l…

🎉 New blog post on a better (visual!) intuition for information theoretic quantities (eg entropy and mutual information) 🎉 🔥 Lots of visualisations 🔥 👉 Based on Yeung's "A new outlook on Shannon's information measures" from 1991 📖 #oldiebutgoldie blackhc.net/blog/2019/better…

Is it common or specific to ML that researchers try to add more maths to their papers and complexify their contributions to get through reviews? It is very frustrating to have to parse complexity to find nuggets of simplicity that might not warrant a paper 🙄

Can't believe I made it in the end 🎊😇 Thanks to everyone at @UniofOxford, @ExeterCollegeOx and @OATML_Oxford for the at times stressful, often beautiful, and always inspiring and memorable time 🙏🫶

The state of conferences. 40 weaknesses with 40 questions. Weak reject, confidence 5. Death by sea lioning openreview.net/forum?id=kDhA…

I haven't done a single novel thing in my PhD. I'm just very lucky that reviewers have no clue about prior art 😅

Why are people excited about this paper ("Neural Networks are Decision Trees", arxiv.org/abs/2210.05189)? TL;DR: The result is obvious and useless by itself. Slightly longer "hot" take below 1/4

The "six Hiroshima bombs equivalent" argument is so dumb. All it does is show that Israel has been using precision ammunition to minimize collateral damage because if they had dropped six Hiroshima bomb equivalents indiscriminately on Gaza it would indeed be a parking lot now

I've passed my viva 🥳 Thanks to my examiners @maosbot & @sirbayes for the discussion and feedback! To @tom_rainforth for mentoring; to @joost_v_amersf, @JishnuMukhoti, @fbickfordsmith & @seb_far for our joint papers; to @yaringal for supervising 👨‍🏫 & @OATML_Oxford for the fun 🎉

My Ph.D. thesis (mostly on active learning and information-theoretic intuitions and approaches related to it) is finally on arXiv 🥳 I'm looking forward to finding and fixing many more typos in the future 😂

Why autonomous weapons are inevitable And what we can still do about it medium.com/@BlackHC/why-auto…

🔥 Has your PyTorch code ever crashed because it ran out-of-memory in CUDA, and you had to fiddle with batch sizes repeatedly? 🔥 What if we could just write code that adapted to the available memory instead of resorting to brittle hand-tuning? 🤯 👉 github.com/BlackHC/toma🤗

Very happy & proud to share some research in Deep Bayesian Active Learning from @yaringal, @joost_v_amersf and me at @OATML_oxford 🎉🎉🎉🤗🤗🤗 oatml.cs.ox.ac.uk/blog/2019/…

arxiv.org/abs/2010.06610 is such an insane paper and idea 🤯

Mandatory ELBO derivation in any lecture series: (I think I finally understand the unnecessarily confusing derivation in the VAE paper 😅)

Whoop whoop, I got my first single-author paper accepted 🎉🎉🎉

What a banger 🤯

Intuition why adding Gaussian noise to parameters is nice for optimization: when we integrate/marginalize over the noise, we convolve/blur the loss surface with a Gaussian kernel -> making it smoother

This is one of the best papers I have read in a while. It contains a crazy amount of insights and ideas 🤯

Pretty amazing: they're lowering PyTorch to JAX using a custom torch tensor type and dispatch overrides (all natively supported by PyTorch). It's both amazing and I imagine painful to debug, but great that it works

Can we say that LLMs can't code bc of this paper. Is it fair? No tool use was allowed at all 🤯 How many people can write correct code like that, without running it once to debug or find typos? Even then, the models are in the top 1.5% of human coders. Bad news, indeed 😬

My dream work-life would be 50% basic research and 50% applied engineering work 🤗 And 50% reading papers and books 😅🫠

Working through my thesis corrections and GitHub Copilot gracefully auto-completed both the sentence and auto-generated a comment from my supervisor 😂

This is too real 😂 I've spent the last days reading through the chapters in PML

Why does Google Doc still not support LaTeX equations? 😭😭😭

TRM also provides an intuition that can be directly applied to reasoning LLMs (and RL): To say, that in the TRM algorithm we can view z as thoughts and y as final output (which is compared to some golden solution during training to compute a reward). The `latent recursion` is a refinement (self-critique) cycle of the thoughts with a final update to the solution. The point of `deep recursion` is that we need not take gradients through all steps, but it is sufficient to focus on the last refinement cycle. Now the intuition is clear that for LLMs, we could repeatedly apply a self-improvement operator. starting from empty thoughts and solution (z, y = \empty) or an initial draft. We then compute rewards using a golden solution towards the proposed solution and only backprop through the last refinement cycle, but this should be sufficient to improve the refinement operation. Deep supervision actually has the benefit that bad reasoning will amplify across so many steps, so the final gradient will provide a stronger signal to reign that in

It's a very UK thing that you press the button and by the time it switches to green, there is no more traffic anyway 😂

Different PyTorch versions changing the significance of my results was the world I always dreamt of living in 🙈😅

Singapore is the most impressive city I've ever visited. Way more than NYC or SF. I'm so happy that ICLR (& AABI) decided on Singapore, and I hope we can avoid the US for a bit 🙏 big thanks to my friends for showing me around and to all the awesome people I met and talked to 😊

My current belief is that AI capabilities are more strongly correlated with compute availability than with research itself. Is this wrong?

Why are the NeurIPS reviews released on a Friday and then people are only given 3.5 working days to write rebuttals? Do we care about work-life balance at all? 🤔🐹

Isn't this the whole OpenAI Zürich office? 😂

Several recent papers connect the generalization error of a model to the model’s prediction disagreement. And oh no, I've taken a look at one of them, an ICLR 2022 spotlight, in more detail 🔥 And published my thoughts in TMLR 🥳 1/16 openreview.net/forum?id=oRP8…

My personal bitter lesson is that my random hot take tweets get more views than any of my papers or blog posts that I spend months on 😅🫠

After PhD: would I do it again? No not like that. Was it worth it? No not really. Would I recommend it? YMMV but prob not 😅

It also appears in machine learning a lot, just not sure if anyone has specifically recognized as such. Eg. negative entropy is the log product integral over density using the density measure. It also appears in PAC-Bayes equations a lot

The most effective way to achieve better performance is through pre-training of RL. This unlocks a lot of high-quality data. Right now, pretraining on graduate physics or maths texts is allowed the same compute as text with low information density. The model cannot predict those tokens well right away. Allowing additional thinking tokens during pretraining enables the model to extract a lot more signal from such data

I read "Bias/Variance is not the same as Approximation/Estimation" which is a really nice read on what the title says. There is still so much to be learnt from some of these "simple" equations and decompositions openreview.net/forum?id=4TnF…

Yep, totally support this. Why would anyone not support this?

1/ I just read the fascinating GaLore paper on memory-efficient LLM training using gradient low-rank projection. Kudos to the authors for this insightful work! My TL;DR and some thoughts below (as a little paper review) 🧵

Re memory layers: Shouldn't memory layers include a sink (ie. one memory slot with score 0 which is always included that can be used as none-of-the-above)? Performing softmax on the top-k means that the memory layer can never defer, so you will always grab some random memory locations even if they're not very relevant

🎉🎉Happy & proud to share some research into Information Bottlenecks from @yaringal, @clarelyle and me at @OATML_oxford 🎉🎉 We provide intuition and practical IB objectives for modern DNN architectures, like ResNets. Check it out on arXiv 👉arxiv.org/abs/2003.12537

Yet again feeling slightly overwhelmed by the speed of progress in ML. Kinda wish I could go back in time a few years and re-prioritize certain studies ^^'

For what it's worth, I know lots of folks who essentially work as research scientists without PhD, and reflecting on my own journey, I could have done as much research (if not more with more compute) if I hadn't left to do a PhD in academia ^^' Thad said, doing a PhD has its own perks and benefits ofc

I'm sorry for any hurt feelings for calling NeurIPS PCs clowns and pointing out an apparent domain conflict of interest. I didn't mean any PC individually or personally, but the organization and its (lack of) processes. Sadly, my sentiment was warranted - even if the phrasing might not be to everyone's liking. So, to avoid further tone policing and suggestions to remove the tweet, let me rephrase it to remove any personal mention: NeurIPS is a clown show ("a comically shambolic state of affairs"), and I'm disappointed by the unprofessional official response and the lack of seriousness at the biggest ML conference. I wonder if my takeaway and recommendations should be to take such papers as the role models they are presented as. The incentives are clear: Publish in an area with much hype and fewer knowledgable reviewers; do not worry about attribution; put all of your related work section in the appendix; if someone complains informally, always agree but never act or only to the bare minimum; always wait until after the conference (or at least after the decision notice) to address anything; and if worst comes to worst and someone complains formally, fear not because there are no good processes anyway - everybody knows each other, and there are too many other incentives for anyone to be a party pooper. The ML community has neither teeth nor an appetite for academic integrity. There are too many things that taste sweeter.

If you're not interested in your PhD topic and projects at all, what should you do?

It's about the University failing the student by not raising issues earlier and other academics disagreeing with that assessment in the first place. You are not supposed to fail a student at confirmation for things that should have been raised at the transfer of status viva

Somebody please explain high-dimensional embeddings to me and what things look like in those spaces 😅

I might be finishing my PhD within the next half year. What does one do after? I guess the time to apply for stuff was a few months ago 😅

Publishing a paper at ICML costs 650 USD, which is cheaper than some journals, but it cannot beat TMLR, which is free and comes with higher-quality reviews 🥳

Due to popular demand: for example, revising basic maths some more again (LA, calculus), some more ML basics (kernel methods, SVMs, etc), old DL papers. But also orthogonally: for various projects, spend more time on lit reviews/prior art and playing around with the baselines...

Finally catching up on MoE with @finbarrtimbers's great posts on this (link below). My thoughts MoE and objectives: 1. Instead of the weirdly unprincipled additional losses, one can simply maximize the mutual information I[E;T], where E is the expert index and T is the token: I[E; T] = H[E] - H[E | T], so maximizing the mutual info maximizes the entropy of expert selection without knowing the token H[E], ie all experts selected uniformly across all data, while it minimizes the entropy knowing the token H[E | T], ie as one-hot as possible for a given token, equivalent to K=1. 2. The softmax(top-K of router logits + normal noise) is almost equiv a double softmax: Taking the top-K (router logits + Gumbel noise) is equivalent to samping from softmax(router logits) k times w/o replacement. Applying the softmax to those samples simply distributes the credit accordingly between the top-K chosen experts. A potentially cleaner formulation would simply always use a full mixture and only look at the top-K sampling approach etc as performance optimizations. 3. The "router Z-loss" seems overcomplicated. Z seems to stand for the partition constant of the induced categorical distribution of the logits. The Z loss does not affect router predictions as it affects all expert logits in the router equally, and it is motivated by numerical stability. Instead of regularizing with Z loss explicitly as a loss, one could also simply adapt the bias of the router network and shift it by the mean logit activations of a training batch. Same effect and no loss needed. 4. Why do we use MoE only for FFNs and not for attention? MoE for QKV or at least for the Q matrices would seem quite valuable to make attention token-specific and either save FLOPs or get better attention for same FLOPs. Mixture of Depths seems to look at that finally.

Had to check: it is real but book-bait... uses "a one-sided generalized derivative called a subdifferential" instead of derivatives link.springer.com/book/10.10…

“Stochastic Batch Acquisition: A Simple Baseline for Deep Active Learning” has been published @TmlrOrg 🎉 My last paper with @OATML_Oxford and the amazing @seb_far*, @parmi_atg, @anndvision, @fbranchaud1, and @yaringal Some details and a blog post below:

So riding a bicycle requires petrol bc your Amazon parcels are delivered by truck? Same logic 🙄 I wouldn't trust a founder making such an argument with anything

ML design interview question @ @Waymo: they log a lot of online telemetry data, too much to transmit it all. only some of that data is interesting when run in simulation later because it might lead to divergent and wrong behaviour by agents in the simulation.

arxiv.org/pdf/1905.12957.pdf “Neural Entropic Estimation: A faster path to mutual information estimation” 👈 this paper is beautiful It derives the Donsker-Varadhan representation en passant using simple straightforward steps to get to MINE’s estimator...

A small info-theory thread (or at least food for thought): Why is the Bayesian Model Average the best choice? Really why? I'll go through a naive argument (anyone has better references?), simple lower-bounds and decompositions, and pitch a "reverse mutual information" 1/15

We are though? The overall trajectory is still one of continued global mass extinction, and this is putting lipstick on a pig

I'm bored, so let's examine how the two losses are maybe similar: d_i = ||w_i - x||^2 = -2 w_i . x + ||w_i||^2 + ||x||^2 -log d_i \approx 1 - d_i So then: 1/d_i = exp(-log d_i) \approx exp(2 w_i . x - ||w_i||^2 - ||x||^2 + 1) = exp(2 w_i . x + C_i) where C_i = - ||w_i||^2 - ||x||^2 + 1 The harmonic loss probability looks approximately like a softmax: 1/d_i / (sum_j 1/d_j) \approx exp(2 w_i . x + C_i) / (sum_j exp(2 w_jY . x + C_j)) Softmax is invariant to shifts, so what if the C_j are approximately constant? Empirically, "Understanding Softmax Confidence and Uncertainty" by Pearce et al. (2021) [1] argues that all w_i are about the same magnitude for optimal decision boundaries when trained using CE loss (so assuming we train the model for long enough). Then, ||w_i|| \approx const, and all C_i are approx const for a given x. Thus : exp(2 w_i . x + C_i) / (sum_j exp(2 w_j . x + C_j)) \approx exp(2 w_i . x + C_i) / (sum_j exp(2 w_j . x + C_i)) = exp(2 w_i . x) / (sum_j exp(2 w_j . x)) This is the same as the regular softmax in the cross-entropy loss with temperature 1/2, so slightly sharper. What does this tell us? Not much, but I was bored, and this was fun 😊 --- [1] arxiv.org/abs/2106.04972

Finally, another paper summary from me for the amazing "Limitations of the Empirical Fisher Approximation for Natural Gradient Descent" by Kunstner, @lukas_balles & @PhilippHennig5 notion.so/Limitations-of-the…

Very happy to announce that my reproduction "Does ‘Deep Learning on a Data Diet’ reproduce? Overall yes, but GraNd at Initialization does not" has been published in @TmlrOrg 🥳 1/5

Can't wait for the series about it: Game of Flops

Kinda depressing when you keep getting rejection emails for internships 10 months later without even being asked for an interview 😐

Weirdly, I'm rooting for China somehow. How did that happen?

A lot of finance bros are discovering that they wanted to do AI research all along right now

Nah it's a human em-dash bc real em-dashes don't have whitespace around them, and an LLM would know better