RanW

Write about machine learning and cognitive science topics at

[RanW @_RanW_]

@pietromazzaglia @1x_tech Welcome to the bay Pietro!

[RanW @_RanW_]

@natolambert Olympic park good place to run!

[Conor Heins @conorheins]

pymdp 1.0.0 is here: batched, autodifferentiable, JIT-compiled active inference in JAX: https://t.co/Hhzsh1wOv5
This release brings:
GPU/TPU-ready active inference
autodiff through inference, planning and learning
easy parallelization and batching with vmap()

[RanW @_RanW_]

@agarwl_ Actually R1 optimality I think

[RanW @_RanW_]

@agarwl_ Use R2 as reward shaping if multi step? Guaranteed R1 improvement I think

[RanW @_RanW_]

@Waymo into 2026💨

[RanW @_RanW_]

@SarahChieng Honestly the other run this morning was a bit too fast 😅 slow is great

[RanW @_RanW_]

@grx_xce Hi, can’t attend but is there a way to learn more about this?

[RanW @_RanW_]

Sometimes it’s very useful to write down the Bayes net/factor graph/causal DAG of your env.

Helped me quite a lot e.g. studying Alchemy: https://t.co/2WKAGmxhif

[Pablo Samuel Castro @pcastr]

🚨The Formalism-Implementation Gap in RL research🚨

Lots of progress in RL research over last 10 years, but too much performance-driven => overfitting to benchmarks (like the ALE).

1⃣ Let's advance science of RL
2⃣ Let's be explicit about how benchmarks map to formalism

1/X

[Harshit Sikchi @harshit_sikchi]

At @RLBRew_RLC today we are presenting 2 works on unsupervised RL and 1 work on inverse RL. Stop by the poster session to learn more! Details below:

[Alec Tschantz @a_tschantz]

New paper from @VERSESAI - AXIOM is a world model that learns to play pixel-based arcade games in minutes.

Preprint: https://t.co/DnqwZDaN6l
Blog: https://t.co/bIsschojXA
Code: https://t.co/nxQKlm8swF
🧵

[RanW @_RanW_]

https://t.co/ezFJy4mW9I

[RanW @_RanW_]

New post studying the empowerment objective for the assistance game in human-AI collaboration. What is empowerment optimizing? Is it aligned with human preference? What's the ultimate objective for human-AI collaboration? 👇

[RanW @_RanW_]

We also found some useful implementation tricks and tips and observations along the way. These details are documented in this blog (https://t.co/wXXyakWhhP).

[RanW @_RanW_]

Hi imitation learning friends, I am excited to introduce CleanIL (https://t.co/jacC3COKXZ), a repo of high quality single-file implementations of imitation learning and inverse RL algos inspired by CleanRL and built on @torchrl1.

[RanW @_RanW_]

CleanIL aims to address this by gathering SOTA algos scattered all over the internet into a single repo. We implemented 7 algos as a starting point. Future plans are outlined in this blog post (https://t.co/Bsunt8kwtB) along with interesting use cases of IL and IRL.

[RanW @_RanW_]

@TimDarcet @jon_barron It’s pretty much mixture density network https://t.co/y6JfuVHERx

And ppl have tried EM Gaussian splatting

[RanW @_RanW_]

@Laz4rz Not as complicated as your lecture but hopefully help with intuition: https://t.co/A0hdW3Gtfl

[RanW @_RanW_]

@RichardMCNgo If you still have some lingering doubts, these might be helpful:

https://t.co/jGdqhkaeYe

https://t.co/MH4LYJX5eJ