Running artificial curiosity sweeps on CSCS

Do not throw away your value network, use it to score paths:

https://arxiv.org/abs/2309.15028

ACL 2023 rolling review: https://www.aclweb.org/portal/content/submission-dates-and-process-eaclnaacl-and-acl-2024

You can submit to one conference (EACL or NAACL), but commit to another (e.g. ACL) after receiving reviews with an option to resubmit.

Dates:

15 Oct 2023: October ARR Cycle - EACL submission deadline
15 Dec 2023: ARR reviews & meta-reviews available to authors of October cycle
15 Dec 2023: December ARR Cycle - NAACL submission deadline
20 Dec 2023: EACL commitment deadline
15 Jan 2024: EACL decisions available
15 Feb 2024: ARR reviews & meta-reviews available to authors of December cycle
15 Feb 2024: February ARR Cycle - ACL submission deadline
20 Feb 2024: NAACL commitment deadline
15 Mar 2024: NAACL decisions available
15 Apr 2024: ARR reviews & meta-reviews available to authors of February cycle
20 Apr 2024: ACL commitment deadline
15 May 2024: ACL decisions available

Bayesian Flow Networks official code has been released. https://github.com/nnaisense/bayesian-flow-networks

How to count experience? This post introduces a counter for embeddings: https://proger.github.io/posts/counter/counter.html

I made a post on neural circuits, in the context of the gradient vanishing problem, linearity and autoregression:

https://proger.github.io/posts/neural-circuits/recurrent.html

Deep Reinforcement Learning has been used to develop champion-level FPV drone racers, with a paper in Nature appearing 10 years after DQN was developed to beat humans in Atari. Key success lies in development of high quality simulations and sim2real policy transfer. RL is no longer about games.

One of the authors, Vladlen Koltun was giving a seminar in MIT: https://www.youtube.com/watch?v=vNFTcD3QMn0

Here’s the system demonstration of Swift https://www.youtube.com/watch?v=fBiataDpGIo

Meme moment, my pure torch linearized LSTM implementation just beat GPT in throughput in the small (110M) setting. This is the fastest LSTM implementation to date. More experiments soon.

zero-shot learning by searching for support vectors for generative tasks

https://openreview.net/forum?id=QBlegfNZNE

irfft(rfft(ones(4), n=4*2) * rfft(arange(1,5), n=4*2))[:4]
= cumsum(arange(1,5),-1)

Raining tokens 2 https://x.com/statmlpapers/status/1717353357772300342

Scalable instance-level meta-learning by looking at every token having its own reconstruction task

https://arxiv.org/abs/2310.13807

https://twitter.com/srush_nlp/status/1720113524121235577

Fine grained control of KV caching is one of the reasons I work on haloop.

Install today: https://www.youtube.com/watch?v=2-G5bomAkfs

Prof reminds us to add gradient noise and decay learning rate. Probabilistic AI 2023 is deep learning lab in math speak

Don’t regulate AI, regulate uses of AI

https://laion.ai/blog/transparent-ai/

torch.export can unroll for loops that depend on the time dimension of the input. Sequential RNNs go brrrr!

New work from IDSIA. Transformer's largest bottleneck is a feedforward MLP block that expands hidden dimension four times and shrinks it back.

Instead of running all parts of this network on every request, sparse gating decides what subnetwork to run depending on the input. This is straightforward at inference time but hard to backpropagate at training time.

Sparsely-gated Mixtures of Experts date back to Ivakhnenko and Lapa (see Section 4) , and now feature an open source implementation from Robert.

https://arxiv.org/abs/2310.10837

Btw does anybody know how to use Triton kernels with torch.compile without graph breaks?

ICLR 2024 has a track that accepts anonymized blog posts. Submissions are now open:

https://x.com/fpedregosa/status/1722197379418665247?