ETH Zurich University of Zurich ETH AI Center
Elia Torre - 15.04.2026 1

Diffusion RNNs

Modeling naturalistic behaviors with noise

Contents

  1. 01Looking behind
  2. 02Noise-diffusion RNN dynamics
  3. 03Discovering structure in natural behavior
  4. 04Looking ahead

Naturalistic behaviors

  • Complex
  • Internally- and externally- driven
  • Probabilistic

Computation as dynamics

Computation as dynamics (modeled with RNNs)

Winner-take-all competition
Douglas et al, 1995 Wang, 2002 Machens et al, 2005
Pattern generation
Churchland et al, 2012 Elsayed et al, 2016
Evidence integration
Mante et al, 2013

Current models of behavior

Wiltschko et al, 2015 Weinreb et al, 2024

Can the continuous state space of RNNs give rise to discrete, probabilistic behavioral structure?

Training RNNs to approximate Hidden Markov Models

Torre et al, 2025

Double-well potential hypothesis

Single stable fixed point

Clusters, kick-zones and transitions

Clusters, kick-zones and transitions

Clusters, kick-zones and transitions

Clusters, kick-zones and transitions

Cluster
Transition

Clusters, kick-zones and transitions

What happens in the
“Kick-Zone”?

Self-induced stochastic resonance

A compositional dynamical primitive

= Single fixed point + Noise effect + Cluster & Transition + Kick–zone

A compositional dynamical primitive

= Single fixed point + Noise effect + Cluster & Transition + Kick–zone

Latent states of natural behavior

Calhoun et al, 2019

Clusters and feature gradients

Take-Away

  • RNNs can emulate HMMs–like behavior.
  • Discrete behavior emerges from continuous dynamics using a compositional motif.
  • These dynamical properties extend to biological data, where RNNs uncover a richer latent structure than HMMs.

Take-Away

  • RNNs can emulate HMMs–like behavior.
  • Discrete behavior emerges from continuous dynamics using a compositional motif.
  • These dynamical properties extend to biological data, where RNNs uncover a richer latent structure than HMMs.

Take-Away

  • RNNs can emulate HMMs–like behavior.
  • Discrete behavior emerges from continuous dynamics using a compositional motif.
  • These dynamical properties extend to biological data, where RNNs uncover a richer latent structure than HMMs.

Thank You!

  • Liyuan Li
  • Michele Viscione
  • Lucas Pompe
  • Renate Krause
  • Goncalo Guiomar
  • Valerio Mante

Master Theses & Collabs

  • Beyond the flybio / ML Apply Diffusion RNNs to mouse, bird, primate data; surface latent structure HMMs miss.
  • Pushing the frameworkCS / ML Scale to richer behaviors via new losses, architectures, biological constraints.

Come talk to me!

Provide a feedback!

Thank You!

  • Liyuan Li
  • Michele Viscione
  • Lucas Pompe
  • Renate Krause
  • Goncalo Guiomar
  • Valerio Mante

Master Theses & Collabs

  • Beyond the flybio / ML Apply Diffusion RNNs to mouse, bird, primate data; surface latent structure HMMs miss.
  • Pushing the frameworkCS / ML Scale to richer behaviors via new losses, architectures, biological constraints.

Come talk to me!

Provide a feedback!

Thank You!

  • Liyuan Li
  • Michele Viscione
  • Lucas Pompe
  • Renate Krause
  • Goncalo Guiomar
  • Valerio Mante

Master Theses & Collabs

  • Beyond the flybio / ML Apply Diffusion RNNs to mouse, bird, primate data; surface latent structure HMMs miss.
  • Pushing the frameworkCS / ML Scale to richer behaviors via new losses, architectures, biological constraints.

Come talk to me!

Provide a feedback!