iiinova — where your digital self keeps evolving

Louis Pezon @lpezon.bsky.social · 28/09/2026

How can RNNs re-use the same dynamics across different tasks to implement compositional computation? What are the consequences for network connectivity & activity? If you're interested, come and find out at the Bernstein Conference, poster session III-#48!

0182

Reposted by Louis Pezon

GerstnerLab @gerstnerlab.bsky.social · 29/06/2026

How can the brain learn the hidden hierarchical structure from high dimensional data? In our latest work, we use synthetic datasets to analyze two classes of bio-plausible learning rules: variants of Direct Feedback Alignment, and local self-supervised learning. We find only the latter succeeds.

12612

Louis Pezon @lpezon.bsky.social · 30/06/2026

Finally, how do 𝑠ℎ𝑎𝑟𝑒𝑑 latent components enable 𝑡𝑎𝑠𝑘-𝑑𝑒𝑝𝑒𝑛𝑑𝑒𝑛𝑡 computation? Our framework explicitly identifies several potential loci at which task-dependence can be resolved. This yields a qualitative, interpretable description of compositional computation in RNNs.

050

Louis Pezon @lpezon.bsky.social · 30/06/2026

We derive consequences of such shared 𝑙𝑎𝑡𝑒𝑛𝑡 components: (i) They are compatible with task-dependent activity, observed consistently in brains and RNNs. (ii) They constrain connectivity statistics and neural representations, and are detectable via specific experiments.

160

Louis Pezon @lpezon.bsky.social · 30/06/2026

We develop a theory of compositional computation in low-rank RNNs. Low-rank nets perform computation via low-dimensional latent dynamics. We propose that “shared components” exist at this level — more precisely, that networks re-use 𝑎𝑢𝑡𝑜𝑛𝑜𝑚𝑜𝑢𝑠 𝑙𝑎𝑡𝑒𝑛𝑡 𝑑𝑦𝑛𝑎𝑚𝑖𝑐𝑠 across tasks.

150

Louis Pezon @lpezon.bsky.social · 30/06/2026

Both brains and RNNs can re-use components of computation across similar tasks or contexts. But what exactly are those “shared components”? How can they be used to solve several tasks? We address these questions in a new preprint with @avm.bsky.social! Link: www.biorxiv.org/content/10.6...

biorxiv.org

Interpretable compositional computation with recurrent neural networks

Flexible cognition utilizes reusable components to enable rapid adaptation of behavior to different contexts or tasks. Analysis of artificial neural networks trained on multiple tasks suggested that t...

16425

Reposted by Louis Pezon

Flavio Martinelli @flavioh.bsky.social · 10/06/2026

NEW PAPER. Why do larger networks train better? "Because they contain more candidate *sub*networks that can learn the task" → lottery tickets This popular explanation uses an appealing but misleading metaphor🧵 We propose an intuitive alternative grounded in theory: escape dimensions

519952

Louis Pezon @lpezon.bsky.social · 13/03/2026

Presenting a poster tomorrow at Cosyne 26: [3-033] Compositional computation via shared latent dynamics in low-rank RNNs. With @avm.bsky.social, we explore how RNNs can re-use the same dynamics across different tasks, and what it implies for their connectivity and neural activity.

0102

Reposted by Louis Pezon

Alexander van Meegen @avm.bsky.social · 12/03/2026

Travelling to COSYNE seems to be the perfect opportunity to announce that I started my own lab at RWTH Aachen University earlier this year, funded by NRW's Ministry of Culture and Science through its Return Program. If you are at COSYNE and want to chat please reach out!

5335

Louis Pezon @lpezon.bsky.social · 06/03/2026

[About the theory] To tackle this problem, we developed a unifying framework for the theory of neural fields and low-rank RNNs. We show that low-rank nets rely on an implicit spatial structure, and conversely, that neural fields obey latent dynamics akin to low-rank nets.

060

Louis Pezon @lpezon.bsky.social · 06/03/2026

3. We find that the circuit structure – can impose symmetries on the network's low-dim. dynamics – constrains the topology of the set of all single-neuron responses. From the second point, we pinpoint topological features of neural activity relevant for comparing it with models.

161

Louis Pezon @lpezon.bsky.social · 06/03/2026

2. We find that “function does not determine form”: the same computation (i.e., the same low-dim. dynamics) can be implemented by networks with different circuit structures. Yet, circuit structure imposes subtle constraints on neural activity (see next).

150

Louis Pezon @lpezon.bsky.social · 06/03/2026

1. Neural activity is often analyzed in two ways: – low-dimensional dynamics in neural manifolds reflect computations – single-neuron response properties (e.g., tuning, selectivity) can describe how the neural population is organized. But it's not clear how this relates to network connectivity.

120

Louis Pezon @lpezon.bsky.social · 06/03/2026

Excited to share our new paper to be published in Neuron! With Valentin Schmutz @bio-emergent.bsky.social and Wulfram Gerstner @gerstnerlab.bsky.social, we explore how circuit structure in RNNs shapes network computation and single-neuron responses. www.sciencedirect.com/science/arti...

sciencedirect.com

Linking neural manifolds to circuit structure in recurrent networks

Dimensionality reduction methods are widely used in neuroscience to investigate two complementary aspects of neural activity: the distribution of sing…

14211