Reinforcement learning followed by mechanistic interpretability on mice modulated by ketamine

Why and how to initialize neural networks? We would like them to undergo "nontrivial" updates when going through stochastic gradient descent — a regime that is also called "feature learning".

This regime asks for the the norm of the hidden features *and their gradient* scaling like the square root of the width of the feature.

You can achieve this desiderata by scaling your initialization and learning rate to preserve certain properties of the spectral norm of your weights and gradients.

By staying in the feature learning regime, you also get predictable hyperparameter transfer from small-scale development proxy models to large-scale expensive training runs — something you don't get from PyTorch defaults.


https://arxiv.org/abs/2310.17813

Jan Leike authored his PhD thesis on Nonparametric General Reinforcement Learning showing that Hutter's incomputable but optimal AGI, AIXI, collapses under degenerate choices of priors and can't be optimally approximated using finite computation. He then shows that AIXI can be approximated epsilon-optimally and provides alternatives to asymptotically optimal learning in stochastic environments based on Thompson sampling.

https://jan.leike.name

I'm finally citing Schlesinger!

https://sites.google.com/view/ngsmworkshop

Call For Papers: ICML 2024 Workshop on Next Generation of Sequence Modeling Architectures

Submission Deadline: May 31, 2024 (Anywhere on Earth)
Acceptance Notification: June 17, 2024 (Anywhere on Earth)

I am happy to be serving as a reviewer for this workshop. Looking forward to learning new insights into sequence models from you.

My best science joke so far

tired: github streaks
wired: wandb streaks

New challenge for signal recognition: the bandwidth has increased

https://content.neuralink.com/compression-challenge/README.html

Automatic learning rate transfer across sizes is now easier to use: https://github.com/jxbz/modula/tree/main

Manifest AI is working on linear transformers and context scaling. At the end of this article authors discuss what is possible when you push the current context size limits — at billions of tokens you won't need finetuning any more — you'll just be able to push your entire dataset into the context window.

Currently open source LMs are at thousands of tokens, industrial grade LMs are at the millions of tokens — there's a lot of work left to push this frontier. In transformers we are simply concatenating token embeddings to the memory, and we will need some automatic compression to get past this.

https://manifestai.com/articles/compute-optimal-context-size/

4o has become so good at math that it can analyze recurrences for me (and tolerate my typos)

Doing symbolic differentiation with loops is a piece of cake, I don't have to explain what a "backwards pass" is: https://chatgpt.com/share/858b2882-9d29-442e-a0cb-7e3afb24abab

reproducing gpt-2 small is now about 10x cheaper than last year:

https://x.com/karpathy/status/1795484547267834137/photo/2

OMNI reboot as OMNI-EPIC: ask ChatGPT to generate environments for your agent that are progressively harder to solve!

https://x.com/jeffclune/status/1795787632435212732

The recent CoPE paper made it all click. Gated Linear RNNs are transformers with token-dependent masks without softmax. If you would like to know how to calculate the attention mask, read my upcoming thesis :D

Spoiler: it’s dynamic programming!

Dynamic programming is also used to find optimal tensor contraction paths, expoiting associativity of tensor products

The screenshot uses opt_einsum

Тимофій Милованов дає приклади байесівського інференсу:
1. дати prior через промпт моделі (спитати що таке organizational culture principles by Ed Schade)
2. дати likelihood моделі (зкинути датасет та спитати identify cultural misalignments)
3. отримати posterior — обговорити результат

https://youtu.be/LTpWpadoT_U

A few of recipes for cleaning pretraining data are popping up:

- FineWeb-Edu, maximizes scores on downstream educational benchmarks

https://x.com/gui_penedo/status/1797173053123916036

The ablations are 1.82B training runs for 30B tokens (almost same as uk4b-large in data, >2x the size) — ~1 GPU-month per one ablation!

> Our ablation models were trained using nanotron. Our "ablation models" have 1.82B parameters (including embeddings), used the Llama architecture with a 2048 sequence length, a global batch size of ~2 million tokens, and the GPT2 tokenizer. For most ablations we trained on ~28B tokens (roughly the Chinchilla
optimal training size for this model size). To confirm relative performance improvements after each step of filtering we conducted longer training runs on 350 billion tokens as mentioned further below


And similar work from the FLAN collection author:

https://arxiv.org/abs/2305.13169

https://www.youtube.com/watch?v=i2Mp_Bc14WI

Since FPV pretraining is actually mainstream (= best ICLR honorable mention paper) I'm now obsessed with looking up animal camera videos on YouTube. Do you have a favorite of your own?