Information & Decisions Across Scales: Constraints and Optimality in Neural and Biological Networks

Cells use highly interconnected and noisy biochemical networks to sense and process the information contained in their environment. In order to establish a mechanistic understanding of such processes, both the structure and kinetic parameters of the corresponding network are necessary. However, simultaneous measurements of all relevant parameters are not readily possible. In addition, the stochastic nature of biochemical reactions at the cellular level necessitates probabilistic approaches where intrinsic noise is incorporated. Here, we propose a method for inferring the dynamics of large biochemical networks with a candidate network structure but missing parameters. Starting from a general chemical Langevin eq. and the corresponding Fokker-Planck eq., we derive a set of moment equations that are closed by assuming a Gaussian distribution for all species. A Bayesian inference scheme then enables us to identify the network parameters along with their confidence intervals. Furthermore, using the concept of accessibility from graph theory, we identify inferable parameters for given initial conditions of concentrations. We demonstrate the efficiency of our approach by identifying network parameters in two simple toy models and a single-cell biochemical network exhibiting memory and basic form of learning.

Cells must reliably regulate gene expression despite substantial intracellular fluctuations. Stochasticity in transcription and translation generates high variability in protein concentrations, yet cells achieve robust function through mechanisms that buffer these fluctuations. Biomolecular condensates (Biocondensates) formed by liquid–liquid phase separation provide one such mechanism: by dynamically partitioning molecules between dilute and dense phases, condensates help maintain the free-particle concentration fluctuations within a narrow range even when total molecular copy numbers vary. Experiments and theory have demonstrated noise reduction in single-component condensates under near-equilibrium conditions. However, endogenous condensates are inherently multicomponent and operate far from equilibrium due to molecular turnover, active processes, and compositional regulation. How these features jointly influence noise control remains poorly understood, and a framework that simultaneously accounts for multicomponent thermodynamics and non-equilibrium kinetics is still lacking. In this work, we develop a theoretical and computational approach to quantify fluctuations in out-of-equilibrium multicomponent Biocondensates. We combine condensates thermodynamics with stochastic kinetics and use a system-size–expansion approach to quantify mean concentrations and their fluctuations. This framework lays the groundwork for a unified description of how compositional complexity and non-equilibrium activity shape the noise-buffering capabilities of cellular condensates.

Navigation at the nanoscale, from synthetic nano-robots to living microorganisms, occurs in environments shaped by both thermal fluctuations and structural disorder. Understanding how such disorder reshapes optimal decision-making is essential for applications ranging from targeted drug delivery to atomic-scale fabrication processes [1–3]. We investigate the optimal navigation of a Brownian particle driven by a force that biases its motion in a frozen disordered energy landscape, modeled as a lattice with randomly distributed traps [4]. Using dynamic programming, we compute optimal control policies that minimize the mean first- passagetime to a target site. We then quantify how environmental disorder modifies these policies. We show that the probability of local policy change induced by disorder is directly related to the Kullback–Leibler divergence between the optimal policies distributions in the disordered and homogeneous environments. Remarkably, this probability exhibits a non-monotonic dependence on defect density, reaching a maximum at low defect concentrations. In the fluctuation dominated regime, we derive an analytical expression for the distribution of optimal policies. Furthermore, the location of the maximum occurs at a concentration which scales inversely with trap strength. Importantly, this behavior does not depend on a specific form of the transition rates, and can therefore be extended to a broad class of navigation models, including those relevant to animal and robotic navigation. [1] F. Novotny et al., Chem 6 (2020) 867–884. [2] G. Li et al., PNAS 105 (2008) 18355–18359. [3] P. Leinen et al., Sci. Adv. 6 (2020). [4] K. Bilaï Biloa, O. Pierre-Louis, in preparation (2026).

Retinal ganglion cells (RGCs) are the sole output neurons of the retina, providing a unique system for understanding how sensory perception is transformed into neural representations. RGCs are divided into several functional types that encode complementary features of the visual scene, each organized in mosaic-like patterns that tile the entire visual field. Yet despite decades of study, the number of RGC functional types remains unresolved, in part because robust, precise functional classification at scale is still lacking. Standard functional typing for large-scale recordings relies mainly on responses to the chirp stimulus -a full-field modulation of light intensity over time-. However, full-field stimulation disregards spatial computations that are known to occur in the retina (e.g., center-surround interactions) and often fails to distinguish similarly responding RGC types. To address this issue, we developed a barcode stimulus: a moving pattern of black-and-white stripes spanning multiple spatial frequencies and applied it to ex vivo mouse retinas recorded with a multi-electrode array (MEA). Compared with the chirp, the barcode evoked stronger and more reliable responses across trials (higher SNR and lower CV). Crucially, its spatial richness exposes consistent differences between cells that appear indistinguishable under chirp stimulation. We analyzed barcode responses using a combination of t-SNE (t-distributed stochastic neighbor embedding) and a targeted feature-extraction approach that quantifies each neuron’s response to every stripe in the barcode. Preliminary results indicate that this clustering pipeline organizes cells into broad neighborhoods consistent with canonical response classes (ON/OFF; transient- vs sustained-like). Within each neighborhood, it further resolves finer clusters whose receptive-field locations form mosaic-like patterns. These results were consistent across independent experiments. Together, our results highlight two key points. First, achieving fine resolution of neuronal functional types requires a stimulus that elicits sufficiently rich responses---here, the barcode separated groups that chirp-based procedures could not reliably distinguish. Second, beyond improving cluster resolution, the barcode also drove stronger and more trial-to-trial reliable responses than the chirp, supporting more consistent functional type assignment. We expect that the barcode will provide a fast and reliable approach for assessing functional types in large-scale RGC recordings, a key step toward understanding how visual information is transformed into neural representations.

Organisms engage in navigation to locate critical resources in their environment. Successful navigational strategies require behavioral actions to be coupled to sensory cues. Here we propose an information-theoretic framework that quantifies this coupling using transfer entropy (TE), which measures information flow between sensory inputs and behavioral outputs. Specifically, TE from sensory inputs to behavior defines a "reactive" component of a navigational strategy, in which behavior responds to encountered sensory inputs. Conversely, TE from behavior to sensory signal defines an "active" component of a strategy, in which behavioral outputs generate consecutive sensory inputs in the sensory environment. In this study, we analytically connect these microscopic information flows to macroscopic performance—the steady movement up a sensory gradient—by showing that the geometric mean of active and reactive TE serves as a first-order predictor of a strategy’s performance. This heuristic is derived in a minimal Markov jump model with states defined by local environment (e.g. up, down gradient) and the animal’s action (e.g. run, tumble). We apply the framework to experimentally measured trajectories from bacteria, worms, flies, and machine learning agents navigating sensory landscapes. Across datasets, the geometric mean of the two TE directions reliably predicts navigation efficiency, revealing a common behavioral-environment feedback structure in navigation. Quantifying bidirectional information flows between sensing and action further reveals the diverse strategies underlying bacterial chemotaxis, the spatial dependency of navigation strategy in fly olfactory navigation, the learning dynamics of reinforcement-trained agents, and the effects of memory in navigation.

A classic proxy for altruistic behavior is the donation game, where two individuals must each decide either to be a cooperator, incurring a cost to give a bigger benefit to their partner, or to be a defector and not perform any action (nevertheless receiving a benefit if their partner “cooperates”). Famously, the Nash equilibrium is for both players to defect, even though they would be better off both cooperating. Much work has been done to understand how cooperative behavior can nonetheless be favored. Canonical examples include embedding individuals in a graph (“network reciprocity”) or allowing complex behaviors across iterated interactions (“repeated games”). Recently, this question has been studied when individuals belonging to a collective are subject to a forward-looking institution that can tax and redistribute welfare. The institution’s objective may be wholly selfish, to boost average fitness, or to promote only the well-being of cooperators. In this poster, we study the effects of competing institutions, creating a multi-scale landscape of wealth redistribution and immigration strategies for institutions whose success is invariably reliant on the type and quantity of their respective members, who themselves are attempting to optimize their location via emigration and their personal cooperative strategy.

Habituation is a non-associative form of learning that is characterised by ten hallmarks, which include the typical reversible response decrement upon repetitive encounter of a stimulus as well as frequency and amplitude dependent effects on the strength and timing of habituation. Despite its prevalence in neuronal organisms, habituation has additionally been observed in single cell organisms and individual mammalian cells. This raises the question of how habituation can be implemented on a molecular level, outside the brain. We show that simple molecular networks based on the ubiquitous negative feedback and incoherent feed forward motif robustly exhibit habituation. Furthermore, we find that a combination of two such motifs yields systems that exhibit all hallmarks related to the habituation response to a single stimulus, including effects of the stimulus frequency and amplitude, and repeated cycles of stimulation and recovery. Our models suggest that simple forms of learning, such as habituation, may easily arise from the biochemical networks of individual cells.

L2 regularization drives neural networks (NNs) through transitions between over- and under-parameterized regimes. We show that further transitions of this type exist and are governed by an information hierarchy dictated by data- and task-dependent accuracy regimes. The regime boundaries mani- fest themselves as saddle points in the loss landscape, establishing a direct link between the geometry of the error landscape and the phenomenology of learning. In the linear case this transition can be linked to a symmetry- breaking principle, offering a precise statistical-mechanics analogue, and the L2 regularizer serves not merely as a regularization tool but as a controlled probe of the basin structure of the loss landscape. These results place the phenomenology of regularized learning on a firm geometric foundation. To make the underlying mechanism fully tractable, we derive exact analytical expressions for the critical regularization strengths at transition points using a linear toy model, which serves as an exactly solvable reference point for the theory. Ibrahim Talha Ersoy, Björn Ladewig, and Karoline Wiesner

The study of biological information processing begins with the fundamental concept of internal representation, introduced in cognitive psychology to describe how brains encode and process external data that subsequently shape animal behavior. The abstract representation created compresses the most critical environmental features without losing essential information. While the mechanisms of filtering and representation are most easily observed within the complex nervous systems of multicellular organisms, comparable forms of environmentally-driven learning have been documented in simpler biological systems that lack neural structures, such as bacteria, which is going to be our focus of study. This principle was formalized by the Information Bottleneck (IB) method [N. Tishby 2000], which aims to find a compressed representation of an input variable that preserves as much information as possible about the corresponding output variable. Mathematically, the goal of IB is to identify the minimal sufficient statistics of $X$ with respect to $Y$, that is, a function $T=f(X)$ that maximizes the compression of $X$ while retaining the maximum information about $Y$. To achieve this, the IB method proposes minimizing the functional $$I(X;T)-\beta I(T;Y),$$ where $I(X;T)$ is the mutual information and $T$, which quantifies the amount of information about $X$ contained in $T$. Tanking into account that, in the context of artificial neural networks (ANNs), the compressed variable can be seen as the abstract internal representation housed within the hidden layers [D. E. Rumelhart 1986], IB has been used in the past to understand the learning process between layers in deep neural networks (DNNs) [N. Tishby 2015, 2017]. In a different direction, Recurrent Neural Networks (RNNs)—architectures designed to process temporal sequences— have been shown to be the key structures responsible for maintaining system memory and internal state over time [M. Gabalda-Sagarra 2018, M. S. Vidal-Saez 2025]. In this project we suggest that in order to process information from its environment, bacteria create internal representations through their gene regulatory networks (GRNs). These networks represent the interactions that genes encounter that regulate their expression, which determine the cell state. Bacterial GRNs contain recurrent substructures that can process temporal information from their input and will have specific dynamics. Therefore, we analyzed the updated E. coli GRN [H. Salgado 2023] and classified its nodes as either recurrent—if they participate in at least one cyclic outgoing path—or non-recurrent otherwise. After this classification, we ordered the nodes according to their longest distance from the recurrent hub [M. Gabalda-Sagarra 2018]. By comparing this architecture with randomly generated networks, we confirmed that the number of recurrent nodes is a distinctive feature of biological networks and can be considered a global network motif. Then we took the recurrent nodes from the E. coli GRN and used them as the hidden units of a recurrent neural network (RNN), which we then trained on a delayed AND task, which want to represent how bacteria anticipate a decrease of oxygen after a previous increase of temperature. It is important to note that in our RNN the hidden layer preserves the original GRN connectivity; therefore, during training, we masked the weight matrix to ensure that only biologically valid connections were updated. Finally, we analyzed the information dynamics of the RNN trained on the Delayed AND task using the information bottleneck framework [N. Tishby 2000, 2015, 2017]. We computed the mutual information shared between the input and the predicted output, as well as the mutual information between the predicted output and the expected output throughout the training. What we have observed so far is that, similar to previous work on DNNs [N. Tishby 2015, 2017], our RNN also exhibits two distinct phases during training: an initial fitting phase followed by a compression phase.

Time delay plays a crucial role in many real-world systems and laboratory experiments. It can arise naturally (e.g., due to finite reaction times) or be intentionally introduced (e.g., for chaos control). A key challenge in theoretical modeling and data analysis is determining whether a time delay is present, in what form it appears, and how it influences the system’s dynamics. This involves detecting its presence, characterizing and quantifying its effects, and estimating the delay time. Here, we focus on overdamped stochastic systems and investigate time delay introduced through linear and nonlinear time-delayed feedback forces that depend on a discrete delay time. Such feedback can significantly influence transport properties or give rise to unusual dynamical behaviors, including persistent motion [1] and emergent collective phenomena [2]. By analyzing the power spectral density (PSD), we find that certain features characteristic of analytically tractable linear systems persist in the PSDs of specific nonlinear systems, enabling the inference of time delay. Moreover, even when only a few short temporal trajectories are available, a probing approach combined with deep neural network learning techniques can successfully infer the delay time from stochastic trajectories [3]. References: [1] R. A. Kopp and S. H. L. Klapp, Persistent motion of a Brownian particle subject to repulsive feedback with time delay, Phys. Rev. E 107 (2023) 024611. [2] R. A. Kopp and S. H. L. Klapp, Spontaneous velocity alignment of Brownian particles with feedback-induced propulsion, EPL 143 (2023) 17002. [3] R. A. Kopp, S. H. L. Klapp, and D. Gupta, Inference of time delay in stochastic systems, arXiv:2507.10429 (2025). In press, Physical Review Research.

We present a principle-based theoretical framework for population-dependent cell proliferation that links cellular decision-making to emergent growth laws. The model combines Bayesian adaptation, phenotype regulation, and growth-factor-mediated feedback to describe how sensing limitations shape proliferation dynamics. Using a time-scale separation approach, the coupled dynamics are reduced to an effective population growth law in which inference errors are captured through a mismatch term. This formulation reproduces distinct regimes, including growth enhancement at low density, saturation at high density, and non-monotonic behavior consistent with weak Allee-type effects. The framework provides a compact, interpretable bridge between intracellular information processing and macroscopic population growth behavior

Synthetic immunology aims to develop therapies that leverage the immune system’s ability to reliably distinguish between self and non-self, enabling targeted treatment of cancer and other hard-to-treat diseases. From a physical perspective, this discriminatory capability can be interpreted as immune cells gathering information about their environment. Immune cells actively probe their surroundings by exerting mechanical forces at immunological interfaces. The central question of this work is: What information can a cell extract by pulling on its environment? Mechanical interactions at immunological synapses can therefore be viewed as noisy physical communication channels, where environmental properties are encoded into cellular response states. Motivated by the two-spring model of adhesion mechanics, the immunological interface is represented as a coupled mechanical system. In this picture, one spring characterizes intracellular mechanical properties while the other represents extracellular mechanics. The project analyzes how mechanical fluctuations and thermal noise influence the reliability of information transfer across the interface. In particular, temperature dependence of sensing performance is investigated to understand how physical environment and stochasticity constrain cellular information acquisition.

A central challenge in statistical physics is to formulate a general theory of response for systems operating far from equilibrium. While the fluctuation-dissipation theorem provides a complete framework near equilibrium, it fails for the nonlinear responses to strong perturbations that are common in many physical and biological systems. This talk will address this long-standing gap. The presentation will introduce a rigorous framework for nonlinear response built on a fundamental connection between steady-state responses and mean first-passage time statistics. First, we establish a new, exact identity that universally relates the nonlinear response of any observable under local perturbations to its linear counterpart through a simple, physically meaningful scaling factor. Second, we derive a universal response resolution limit, establisheing a fundamental bound on the change of an observable in terms of its intrinsic fluctuations. The physical significance of these results will be a key focus. We will explore how this inequality serves as a strong-perturbation analogue to the fluctuation-dissipation theorem, setting a fundamental limit on signal-to-noise ratio. The talk will conclude by discussing the direct implications for quantitative biology, showing how these parameter-independent bounds reveal the ultimate physical constraints on biological information processing, for instance, in the context of gene regulation. [1] Bao, R. and Liang, S., 2024. Nonlinear Response Identities and Bounds for Nonequilibrium Steady States. *arXiv preprint arXiv:2412.19602*.

The poster will contain a selection of topics from my doctoral thesis, conducted under the supervision of F. Piazza and J. Hamacek, which was part of X-CROWD, a project aiming at improving the understanding of how macromolecular crowding affects enzymatic activity in the extracellular matrix (ECM). The properties of biochemical reactions that occur in living organisms are influenced by macromolecular crowding in ways that are difficult to predict. However, these reactions are usually studied in dilute solutions. In this project, macromolecular crowding was studied using controlled solutions of the branched polymer dextran. Preliminary characterization of polymer solutions was conducted through rheological and diffusion measurements. Dextran transport properties were found to scale with polymer size, revealing the signature of branching. The rich phenomenology found opens interesting perspectives on the role of dextran as crowding agent and its properties. The enzymatic activity of ECM proteases was tracked through spectrophotometric assays. For this purpose, a theoretical framework for fluorescence detection in non-ideal mixtures was introduced and applied to the full progress curve assays of two key ECM enzymes. In the case of elastase, the degradation of a peptide was found to be enhanced by crowding, revealing an equilibrium constant satisfying the same scaling features as dextran transport properties, suggesting polymer size and topology as tunable parameters in crowding experiments. However, the same result was not observed in the case of a matrix metalloprotease under the same crowding conditions, raising interesting questions about how enzymatic and biochemical specificities combine with solution properties to determine the response to macromolecular crowding. This study lays the groundwork for research on out-of-equilibrium enzymatic reactions in crowded living systems, which is the central topic of the new project $\textit{Dynamics in enzymatic systems out-of-equilibrium with crowding: characterization and inference from observations}$, conceived by D. M. Busiello, on which I am currently working.

Metabolic activity is an indicator of the ‘pace of life’ [1]. Being able to measure metabolic activity not only allows us to understand the physical and biological factors constraining metabolism but also acts as a useful diagnostic tool in biomedicine. One of the direct measures of metabolic output is the heat emanating from an organism (cellular population) but dissipation occurs not only via heat but also chemical fluxes. Further, the metabolic states of cells can exhibit heterogeneity at the single-cell level both in tissues and single cell microbial populations [2]. However, the available techniques to measure metabolic rates of single cellular microbes do not offer insights into population diversity in nutrition uptake [3]. Here, we develop a quantitative measurement technique to isolate and study single cell dynamics without relying on external labels and preprocessing. This method exploits the principle of aqueous flux due to osmotic pressure difference in a unique microfluidics drop- in-drop setup. The rate of nutrition uptake of the microbial cell can directly be inferred from the quantitative changes in the radius of the encapsulating droplet. We first establish this technique rigorously for single celled yeast (S cerevisiae) and then use it to compare the uptake rates of the yeast growing on different carbon sources (sugars). This technique offers opportunity to explore fundamental questions using single cell dynamics such as variation in metabolism at different point in the microbial growth curve, or metabolic heterogeneity for phenotypes at cellular level in microbial colonies. [1] Kleiber, 1947, Body size and metabolic rate. [2] Sieber et al, 2017, The role of metabolic states in development and disease. [3] Boitard, L. et al. Monitoring single-cell bioenergetics via the coarsening of emulsion droplets. Proceedings of the National Academy of Sciences 109, 7181–7186 (2012)

Nonreciprocal interactions that violate Newton's law 'actio=reactio' are ubiquitous in nature and are currently intensively investigated in active matter, chemical reaction networks, population dynamics, and many other fields. The study of the dynamic implications of the 'actio=reactio' symmetry breaking has revealed novel dynamic phases that are relevant to understanding biological processes, such as oscillations, travelling waves, and spiral waves. Despite significant progress in understanding the dynamics of non-reciprocal systems, their thermodynamic understanding remains incomplete. In my talk, I will address this problem. In particular, I will elaborate on the underlying thermodynamic structure that generates such non-reciprocal phenomena and its thermodynamic implications and predictions for non-reciprocal systems that can be experimentally verified.

The ability to store memories and recall them when a similar situation is encountered is central to the survival of many organisms. It is well known that this ability, called auto-associative memory, can be implemented by networks of neurons. Here, we show that it is also possible to construct an auto-associative memory based on interacting genes and proteins that single cells could use to respond to the concentrations, temperatures and mechanical stimuli that they sense. That is done by constructing a blueprint for a gene regulatory network which could in principle be made in a single synthetic cell. The network is based on a synthetic multistable biochemical circuit previously built and analysed by [Zhu et al., 2022]. We demonstrate through simulations that the steady states of this type of circuit can be controlled, either by tuning equilibrium constants according to a simple rule or by adding a slower subsystem that autonomously stores recent memories of expression patterns as steady states and slowly forgets old memories. A mathematical analogy is made between the biochemical circuits and graded response neurons used in theoretical neuroscience, and the recall performance of the networks is analysed as a function of the number of genes.

Here we consider large-scale gene-regulatory networks of non-reciprocal cross-gene interactions, subjected to environmental fluctuations. The network that adapts to the surrounding environment is then selected from those that enable the cell to achieve a specific optimal state. In cellular systems, negative feedback between the fast and the slow components of the gene-regulatory network is the key mechanism to ensure the cell robustly remains in this optimal state, despite stochastic variations. We demonstrate how energy dissipation incurred at multiple scales controls the speed and accuracy of such an adaptation-under-feedback process. Remarkably, we found different scalings of dissipation with the feedback strength required to reach a certain level of accuracy above and below a critical strength, illustrating the central role of non-equilibrium thermodynamics in generating an evolutionary pressure on the double optimisation implemented by regulatory systems: maximisation of performance (or information processing) while simultaneously minimising dissipation.

Thermodynamic uncertainty relations (TURs) provide fundamental constraints on the interplay between power fluctuations, entropy production, and efficiency in overdamped stationary autonomous heat engines. However, their validity in underdamped regimes remains limited and less explored. Here, we analytically and numerically study a physically realizable autonomous heat engine composed of two underdamped continuous degrees of freedom coupled to a two-level system. We show that this nonlinear setup can robustly violate TUR-based trade-offs by exploiting resonant coupling, effectively using one underdamped mode as an internal periodic drive. When this coupling is suppressed, the system recovers TUR-like bounds consistent with overdamped theory. Importantly, we demonstrate that the strongest suppression of current fluctuations occurs in a resonance regime that can be directly inferred from mean current measurements—a quantity typically much easier to access experimentally than fluctuations. Our results reveal new pathways to circumvent classical TUR constraints in underdamped systems and provide practical guidelines for designing efficient, precise microscopic engines and autonomous clocks.

This research investigates the stabilization problem for modelled memristive bidirectional associative memory neural networks (MBAMNNs) with time-varying delays. To be more specific, the structural mechanism of MBAMNNs can be integrated into genetic regulatory networks (GRNs), with memory encoded in gene expression, to assess its potential for delivering a more precise and robust representation. The performance of control signals and actuator fluctuations is evaluated for both MBAMNNs and memristive GRNs (MGRNs), with fault tolerance achieved through state-feedback control design. Based on Lyapunov stability theory and differential inclusion theory, global asymptotic stability criteria are derived in terms of linear matrix inequalities (LMIs). To mitigate the dynamical network’s actuator fluctuations, the fault-tolerant model determines a specific frequency range, enabling effective signal transmission in both known and unknown actuator cases. Finally, the numerical simulation is demonstrated to validate the theoretical findings for optimal stability constraints.

The cytoplasm, nucleus, and lipid membranes of eukaryotic cells contain order thousands of molecules in concentrations ranging from nearly single - copy to millimolar. Biological information is often encoded in a subset of these concentrations, molecular variants that differ along only a few chemical traits such as number of phosphorylated sites, ubiquitin chain length, or binding affinity, rather than in arbitrarily distinct species. The resulting profiles of concentrations across these low-dimensional features carry rich information about cellular state, yet it remains unclear how such molecular distributional codes can be read out without requiring prohibitively many distinct sensors. Collective processes like phase-transitions and assembly-like phenomena are an attractive means to integrate many weak molecular interactions into a collective readout. We query the ability of phase-transitions and molecular binding to discriminate between concentration distributions using mean-field theory, monte-carlo simulations, and experiments in a model system.

Biological neural networks can be remarkably reliable and precise despite their intrinsic noise. This noise is often correlated among neurons within the same layer. The role of this phenomenon, known as noise correlations, remains unclear and is still an active topic of research in neuroscience. This study seeks to determine whether noise correlations are merely a cost of neural connectivity or whether they can, in fact, enhance the information transmitted to downstream neural areas. Experimental evidence indicates that similarly tuned neurons exhibit strong positive noise correlations. A widely accepted theory (Sign Rule) predicts that the observed regime is detrimental to the information encoded in the population activity. Despite its intuitive geometrical interpretation, this theory fails to account for an additional beneficial regime for positive correlations, identified in recent theoretical work. We propose to resolve the inconsistencies between the two theories, providing an enhanced geometrical picture describing the impact of noise correlations on stimulus decoding. We simulated a bar experiment with a two-neuron model with Gaussian noise, in which external stimuli are reconstructed through an optimal Bayesian decoder. We compared the decoding performance of two models, with and without noise correlations. By exploring a range of system parameters, we highlighted the impact of system geometry on decoding performance, identified the SNR-dependent regimes where the two theories are valid, and provided an enhanced geometrical interpretation of the phenomenon.

Water-restricted mice perform an odor-pulse discrimination task in which randomly timed odor puffs are delivered over a fixed stimulus window. Mice then choose between left and right water ports, with reward determined by total number of odor puffs (low → left; high → right). We develop a data-informed reinforcement learning (RL) model to characterize both the decision strategies mice use and how they learn and adapt these strategies over time. We model breathing-induced sensory noise with a probabilistic observation model and formulate the task as a POMDP, which we solve using Q-learning for a two-armed contextual bandit. By fitting parameters that govern exploration and representation of the sensory cues in the brain, we can assess how closely the observed behavior aligns with the optimal policy. The results suggest that sensory noise as well as representation affect decision making and generalization and a minimal tuning allows to explain behavior. Ongoing work extends this framework to characterize learning dynamics and temporal evolution of strategies across training. (working on this project with Francesco Marcolli and Agnese Seminara, mice experiment data are provided by MurthyLab, Harvard University)

Many organisms perceive environmental stimuli and respond to them with dynamic responses. For example, plants respond to light stimuli with a growth response called positive phototropism: plants redirect and grow towards such unidirectional light source, and fluctuate around a particular growth direction even in steady state. We are interested in the quantification of such a response, especially in the context of dynamic mutual information and entropy quantities, provided by the trajectories of these responses. To do so, we discuss analytically tractible expressions and limiting cases for these quantities. Finally, we apply our findings to the plant trajectories and discuss how and when information in dynamic responses can provide an improvement for the quantification of precision, potentially also through decoding maps.

The Ising model is a minimal model of collective binary decision-making, where each site chooses one of two states, considering only local interactions. With spin number conservation, phase separation spatially organizes these local decisions into macroscopic condensates, whose geometries carry imprints of the local rules. Chirality — the breaking of parity symmetry — is a ubiquitous geometric property in nature, from embryonic left-right axis formation to the design of functional metamaterials. Here we propose two minimal chiral Ising models with number conservation to investigate condensate formation, in both equilibrium and active regimes. At equilibrium, we introduce chiral next-nearest-neighbour interactions [1]. This model exhibits multiple novel ground states and chiral condensates emerge with well-defined interfacial structures and orientations. In the active case, we introduce stochastic biased local rotations [2]. At low temperatures, the system undergoes coarsening into chiral faceted condensates, with interfaces aligning at characteristic angles to the lattice axes. These interfaces support persistent edge currents that leave the bulk unchanged and are robust under thermal noise — a signature of unidirectional propagation of local decisions macroscopically. To generalize these findings, we develop a continuum theory by incorporating an active edge current term into "Model B". By constructing an effective interface potential, we explicitly reveal the essential role of active chiral transport in determining condensate geometry. [1] Condensate Formation in a Chiral Lattice Gas, BW, F. Jülicher, P. Pietzonka, New J. Phys. 26, 093031 (2024). [2] Edge Currents Shape Condensates in Chiral Active Matter, BW, P. Pietzonka, F. Jülicher, in preparation (2026).

Many complex systems perform distributed computation through local interactions among their constituents, from ant colonies coordinating foraging to neural networks processing information. Previous studies have shown that granular matter can exhibit rich nonlinear dynamics and memory effects. In this work, we investigate force‑chain networks in granular media as a physical platform for information processing. Strong local interactions and complex contact networks enable time-series input signals to be encoded into high‑dimensional dynamical states that evolve over time. To extract useful outputs, we train a readout with ridge regression, which combines these states into continuous target signals for time‑series prediction. We demonstrate accurate time‑series prediction ability and analyze the memory capacity and stability of the system. The system exhibits decentralized computational properties and maintains performance under perturbations. Our results demonstrate the feasibility of a granular-based distributed physical computing paradigm, in which computation emerges directly from the intrinsic dynamics of granular assemblies, without centralized control or detailed bottom‑up design.