Suri Vaikuntanathan

Work with Agnish, Alexandra, Daiki, and Cal

Excited to share a (long overdue revised version of) work arxiv.org/pdf/2411.07233 : Non-equilibrium active noise enhances generative memory in diffusion models -- TLDR. the ⟨x·η⟩ correlations that drive pressure like effects in active matter help with generative properties and speciation.

We show how it can and also outline a geometric picture that explains ICL even the absence of any attention. ICL emerges almost like a transition when the chemical reaction has enough degrees of freedom to support it.

Super excited to share new work by Cal and Hector arxiv.org/abs/2601.06712 Can In-Context Learning (i.e. a mode of computation typically associated with transformer architectures) emerge in simple chemical reactions ?

This is amazing work !!! @uchichemistry.bsky.social

Neuromodulation-inspired gated associative memory networks:extended memory retrieval and emergent multistability Classical autoassociative memory models have been central to understanding emergent computations in recurrent neural circuits across diverse biological contexts. However, they typically neglect neurom...

Excited to share new work by Daiki and Hector in collaboration with Monika: arxiv.org/abs/2512.13859 We introduce a gating mechanism in classic associative memory models and find capacity is increased far above the Hopfield limit without the usual catastrophic breakdown.

Super happy to be associated with this work !!

Connection between Hebbian Unlearning and Steady States Generated by Nonequilibrium Dynamics Active persistent dynamics are shown to mimic some aspects of Hebbian unlearning --- suggesting that nonequilibrium dynamics can provide an alternative way to improve associative recall in neural netw...

Check out new work by Agnish, Matt and co-workers. Active, persistent dynamics can mimic Hebbian “unlearning,” hinting that nonequilibrium physics offers a new route to sharpen associative recall in neural nets. go.aps.org/49O1yHO @uchichemistry.bsky.social

Yes, Congrats Anna !!!

Super !!!

Preprint 🚨! B cells form localized patterns in the immune synapse when mature, allowing improved affinity discrimination. How? We suggest a new mechanism using dynamic active forces and feedback! Read more @ arxiv.org/abs/2510.18771. Great colab with Shenshen Wang, Tom Chou and Tony Wong (UCLA).

This was so much fun !! And hoping for more @chembiobryan.bsky.social 😀

This paper was a truly @uchichemistry.bsky.social team effort, jointly led by Shannon Lu, Matt Styles, and Frank Gao with me, Aaron Dinner, and @vaikuntsuri.bsky.social – along with a critical support cast. 12/12

Companion to nature.com/articles/s41... [Still under review/revision]

Local imperfect feedback control in non-equilibrium biophysical systems enabled by thermodynamic constraints Understanding how biological systems achieve robust control despite relying on imperfect local information remains a challenging problem. Here, we consider non-equilibrium models which are generically...

Check out arxiv.org/abs/2507.07295 by Cal Floyd where we derived the new non-equilibrium thermodynamic constraint: Intuitively, it says that if we know a system's response at low drive, it will have the same sign at arbitrarily high drive— no matter how close or far from equilibrium !!

Stay tuned for more work on bridging gap between physical principles and computational capabilities of living systems .... @uchichemistry.bsky.social

Importantly, we show how common biological mechanisms—such as promiscuous interactions—can help overcome these fundamental constraints.
This work provides a framework for understanding how biochemical circuits function as classifiers and biological computers.

Limits on the computational expressivity of non-equilibrium biophysical processes - Nature Communications How cells use biophysical processes to interpret complex input signals is not well understood. This study reveals limits to the computational power of generic non-equilibrium systems and shows how the...

Very happy to share work by Cal Floyd now out in Nature Communications nature.com/articles/s41... . We show that the "expressivity" of biophysical decision making circuits is constrained in unexpected ways by non-equilibrium thermodynamics.

This is was a wonderful collaboration ! Hopefully many more ! :)

Learning via mechanosensitivity and activity in cytoskeletal networks In this work we show how a network inspired by a coarse-grained description of actomyosin cytoskeleton can learn - in a contrastive learning framework - from environmental perturbations if it is endow...

check out new work by Deb in collaboration with Martin, @squishycell.bsky.social , Aleks Walczak and Thierry Mora : arxiv.org/abs/2504.15107 o on how simple mechanosensitive agents can enable learning mechanisms

Check out new work by Jordan in collaboration with Aaron Dinner and Petia Vlahovska
! arxiv.org/abs/2503.24120 We consider minimal protocels under non-equilibrium growth conditions and extract low (2) dimensional rules to describe their shapes.

Chemical waves in reaction-diffusion networks of small organic molecules Chemical waves represent one of the fundamental behaviors that emerge in nonlinear, out-of-equilibrium chemical systems. They also play a central role in regulating behaviors and development of biolog...

Very happy to share new work in collaboration with
Sergey Semenov and coworkers. Lissa and Yuqing analyzed the waves in latent space and found a low dimensional representation of the waves. pubs.rsc.org/en/content/a...

Very excited to share a thread about our new preprint! It's a highly collaborative project, w key contributions from @josephlannan.bsky.social (in my lab) @bhamlalab.bsky.social (+ GS Luke Xu), Dinner Lab (+ PD Cal Floyd), Jerry Honts, Marshall lab (+ GS Connie Yan), and Suri Vaikuntanathan 1/13

We need to have a ‘code of conduct’ for reviewing manuscripts, like (1) if you say that something is not novel then provide the reference. (2) If you propose an experiment, state precisely to which claim it is crucial. (3) If you want to say something nasty then sign your name.

Yes , both of you are 🤩

Tailoring interactions between active nematic defects with reinforcement learning Active nematics, formed from a liquid crystalline suspension of active force dipoles, are a paradigmatic active matter system whose study provides insights into how chemical driving produces the cellu...

Check out new work by Cal on using Reinforcement Learning to control the dynamics of defects in an active nematic arxiv.org/abs/2411.09588

(Older post from the other side) Very happy to share work by Cal (w Aaron and
Arvind Murugan
) arxiv.org/abs/2409.05827 ! We show how the expressivity/computational ability of non-equilibrium biological processes may expected to be fundamentally limited.

Score-based generative diffusion with "active" correlated noise sources Diffusion models exhibit robust generative properties by approximating the underlying distribution of a dataset and synthesizing data by sampling from the approximated distribution. In this work, we e...

Check out new work arxiv.org/abs/2411.07233 by Alexandra, Agnish, Aditya and Cal on generative diffusion but with correlated or ``active" noise.

@chembiobryan.bsky.social ok I followed you and @krishnanyamuna.bsky.social