Introducing Latent Program Network (LPN), a new architecture for inductive program synthesis that builds in test-time adaption by learning a latent space that can be used for search 🔎
Inspired by @arcprize 🧩, we designed LPN to tackle out-of-distribution reasoning tasks!
Clem Bonnet
265 posts
Stealth
- Excited for the launch of @ndeainc 🔥 We’re making the bet that merging structured reasoning with deep learning will constitute a new AI breakthrough! Come join us if you want to unlock AI for reasoning tasks and scientific innovation!I'm joining forces with @mikeknoop to start Ndea (@ndeainc), a new AI lab. Our focus: deep learning-guided program synthesis. We're betting on a different path to build AI capable of true invention, adaptation, and innovation.
- Very fortunate to have received the third-place paper award of @arcprize with @MattVMacfarlane!! 🏆 Amazing efforts from @fchollet @mikeknoop and the team to have made working on ARC-AGI such a success 🔥 Excited for 2025!!Today we're announcing the winners of ARC Prize 2024. We're also publishing an extensive technical report on what we learned from the competition (link in the next tweet). The state-of-the-art went from 33% to 55.5%, the largest single-year increase we've seen since 2020. The
- I had a fantastic time discussing latent search with @ecsquendor on MLST! 🙏 We review our work on latent program networks co-authored with @MattVMacfarlane and debate composition, AI creativity, and program synthesis. Full video:We spoke with @ClementBonnet16 at NeurIPS about his extremely innovative approach to the @arcprize using a form of test time inference where you search a latent space of a VAE before making an optimal prediction. @fchollet was so impressed, he hired Clem shortly after! 😃 -
00:00 - Excited to present our two posters on solving CO problems with RL! The inherent method is more general than CO and applies to any ML problems where multiple solutions can be generated at inference time. Come for a chat at posters 1500 and 1421 Awesome colleagues at @instadeepaiTwo papers of my research team are presented at #NeurIPS2023 today. "Combinatorial Optimization with Policy Adaptation using Latent Space Search" (COMPASS 🧭) and "Winner Takes It All: Training Performant RL Populations for Combinatorial Optimization” (Poppy🌹)
GIF - Replying to @ClementBonnet16If you haven't yet, please check out the amazing works from the JAX community. E.g. environments: Brax (@OlivierBachem), Gymnax (@RobertTLange), JaxMARL (@alexrutherford0), Craftax (@mitrma), Pgx (@sotetsuk). And algorithms: PureJaxRL (@_chris_lu_), Stoix/Flashbax (@EdanToledo).
- Integrating deep learning and program synthesis could help create strongly self-consistent world models. This would be extremely valuable for improving how AI systems plan and reason.Replying to @fchollet"system 2 as iterated system 1 with a self-consistency guarantee" is something I considered a few years ago as one of two hypotheses, and it led me to an interesting potential interpretation of consciousness -- consciousness is the consistency guarantee. It explains why all
- We can accelerate #RL research and applications by not only making fast simulators in #JAX (e.g. Jumanji, Brax, Gymnax) but also doing both acting and learning on GPU/TPU, Anakin style [@matteohessel et al.]. Great initiative from @_chris_lu_ for sharing a clean implementation!🔥1/ 🚀 Presenting PureJaxRL: A game-changing approach to Deep Reinforcement Learning! We achieve over 4000x training speedups in RL by vectorizing agent training on GPUs with concise, accessible code. Blog post: chrislu.page/blog/meta-disc… 🧵
- Replying to @Gibbz00 @ndea and @MattVMacfarlaneOne thing is that searching in a continuous latent space is likely faster than a discrete search over tokens, which was our initial motivation. Another is that you can train the model to be good at search (similar to meta-learning), where you can only do RL with tokens.
- 🥳Happy to share our new work published at the MetaLearn workshop at NeurIPS 2022! 💡We propose to learn an alternative value function to fix a bias that arises in meta-gradient reinforcement learning. 📜Paper: arxiv.org/pdf/2211.10550… 💻Code: github.com/instadeepai/ou…Replying to @instadeepaiOur last paper accepted at the Meta-Learning workshop focuses on a bias in current Meta-Gradient RL algorithms (e.g. Bootstrapped Meta-Learning). It fixes it by using an outer value function (6/7) 👨💻 bit.ly/3gEFqXk bit.ly/3EMAjMI
- Highly recommend trying out the quick demo to understand Latent Program Networks. I guarantee it's fun!! Demo HF space:Introducing Latent Program Network (LPN), a new architecture for inductive program synthesis that builds in test-time adaption by learning a latent space that can be used for search 🔎 Inspired by @arcprize 🧩, we designed LPN to tackle out-of-distribution reasoning tasks!
- Replying to @ClementBonnet16Input-output pairs are first encoded into a latent space to serve as an intuitive guess of the program that underlies these pairs. We then refine the latent to be more likely to have generated the pairs. We predict the output of a new input by conditioning on this refined latent.
- Replying to @ClementBonnet16We propose to use gradient ascent (GA) to search through the latent space and show that longer latent optimization leads to higher test-time performance. Gradient ascent can also be used during training, making the latent space optimization aware and leading to better performance










