New preprint! We explore a meta-learning approach to achieving biologically plausible learning in neural networks, using feedback and local plasticity. In our experiments we match (in some cases outperform) gradient descent-based learning. arxiv.org/abs/2006.09549. Thread: [1/n]

Anil Seth on why our senses are fine-tuned for utility, not for ‘reality’ – aeon.co/videos/anil-se… via Aeon Magazine

Judea Pearl In fact, we tackle this problem using selection diagrams here (proceedings.mlr.press/v89/subbaswamy…)! More generally, we discuss the problem of transportability of clinical prediction models as well as types of solutions in this commentary (academic.oup.com/biostatistics/…).

Self-Supervised Policy Adaptation during Deployment (nicklashansen.github.io/PAD/). We propose to adapt RL policy to new environments without any reward, by using self-supervision during deployment. Joint work with Nicklas Hansen, Yu Sun, Pieter Abbeel, Alyosha Efros, Lerrel Pinto (1/3)

Second paper at GECCO 2026 with Danesh Tarapore & D. Bossens: Evolving behavioral descriptors for fast-adaptation in MAP-Elites+BO (IT&E): hal.inria.fr/hal-02555231 [meta-evolution]

I wish I could attend this ICML tutorial today (Monday) on causal reinforcement learning: x.com/eliasbareinboi… -- I've always wondered why RL folks don't sit together with CI folks to create a synergistic ML agenda. Webpage: crl.causalai.net

Language-goal imagination to foster creative exploration in #DeepRL IMAGINE leverages language compositionality and modularity to self-train on out-of-distribution goals We present it @ Language in RL workshop #ICML2020 arxiv.org/pdf/2002.09253… youtube.com/watch?v=0RHx4T… 1/n

Language-Conditioned Goal Generation: a New Approach to Language Grounding for RL Also presented @ the Language in RL workshop #Larel2020 #ICML2020 wth Cédric Ahmed Akakzia M. Chetounai Olivier Sigaud arxiv.org/abs/2006.07043 1/n

Few scientific hypotheses about dreams take dream phenomenology seriously. The Overfitted Brain Hypothesis theorizes that dreams help prevent the brain from fitting too well just to its daily tasks. The very strangeness of dreams assists generalization arxiv.org/abs/2007.09560

Erik Hoel Ryota Kanai This kind of brain/information-focused approach to dreams very valuable, but there are also key explananda for psychology. E.g., dreams about being chased seem to be very common cross-culturally, despite waking chase experiences being thankfully rare! Why? thenextregeneration.wordpress.com/2015/01/05/typ…

Open #postdoc position Inria Flowers team on grounding #AI in the origins of human behavior, bridging #DeepRL and human behavioral #ecology Looking for an interdisciplinary profile #MultiAgent #DeepRL #ComplexSystems #Evodevo jobs.inria.fr/public/classic… (Applying earlier is better)

What next? Logical relationship between action and object and... , common sense, etc? science.thewire.in/the-sciences/o…

If all knowledge comes from sense data, why don't we just gain knowledge the way our ancestors did, by simulating natural selection on our data? Or the way ML folks do, by fitting functions to data? I tried to answer both questions in a new blog post: causality.cs.ucla.edu/blog/index.php…

.Jonathan Brant & I changed the minimal criterion coevolution (MCC) open-ended alg to replace speciation with a simple resource limit with surprising benefit: MCC, now extremely simple, actually works better (& won Best Paper in Complex Systems GECCO 2026!) eplex.cs.ucf.edu/papers/brant_g…

Some brain defects let you perceive social reality in a deeply distorted way while often leaving the rest of your reasoning intact, eg. Capgras illusion: in her last weeks, my grandmother experienced that everyone she knew had been replaced by identically looking impersonators.

An evidence for next gen AI? spectrum.ieee.org/tech-talk/arti…

github.com/silicom-hub/IS… is a new fully versatile and configurable Information System simulator that can be used as a test and demonstration platform, as an educational tool as well as a basis for producing cybersecurity data sets. Feel free to contribute or use it SILICOM

I spent 1000s of hours on competitive programming (proof-link: codeforces.com/profile/rizar). This makes me qualified to comment on #AlphaCode by @DeepMind The result is nice, the benchmark will be useful, some ideas are novel. But human level is still light years away. 1/n

fr.ulule.com/le-prive-5-ans… #ulule #crowdfunding

The great Jay McClelland visits the team Inria Flowers team on 3rd oct ! He will give a talk arguing that modern #AI systems don’t really have thoughts in the same sense that we do, proposing ways forward To attend remotely or in person (10h30am Paris time): 👉forms.gle/79vfsWN2NLmNaj…