Venue: Room 101, Villa Vittoria
The advent of large-scale recordings and detailed connectivity maps, including the recent complete connectome of the fruit fly Drosophila, has provided an unprecedented wealth of data. This progress has been accompanied by rapid development of theoretical frameworks to link neural network structure with dynamics and, ultimately, function and behavior. Each approach addresses specific challenges, such as structural connectivity, functional dynamics, or behavior.
This workshop provides a platform to discuss possible approaches for integrating the fundamental aspects of structure, dynamics, and function. In particular, the workshop aims to identify the properties of neural networks that make them suitable models for neural computation. Additionally, it seeks to define a minimal set of features leading to functional neural networks, thereby addressing conserved evolutionary principles.
To achieve these goals, the workshop is organized into three sessions, followed by a panel discussions:
Structure and Dynamics: How does anatomy constrain information flow? This session features talks from experts on the theoretical analysis of dynamics on networks, the problem of connectivity estimation from activity time-series data, and low-rank networks.
Function and Structure: How does function feed back to structure via learning? Talks will explore connectomics in C. elegans and beyond, generative modeling of brain networks, and self-organizing networks with synaptic plasticity.
Dynamics and Function: How does function emerge from underlying dynamics? Invited speakers will address topics including brain-inspired computing and spiking neuromorphic neural networks, multi-scale modeling of neural networks, dynamics, and computation and dynamical networks supporting learning in the visual system and the cerebellum.
The workshop will conclude with a panel discussion aimed at finding common ground among the three levels: structure, dynamics, and function.
The schedule is tentative.
Abstracts can be found below.
Tuesday, July 8, 2025 (09:00-12:30)
Session: Structure and Dynamics (Chair: Claus C. Hilgetag)
Gorka Zamora-Lopez, UPF, Barcelona, Spain (9:00-9:30): Making (some) sense of the structure-function relationship in brain connectivity
Mariia Popova, UKE, Hamburg, Germany (9:30-10:00): Topological building blocks define attractors in a model of excitable brain dynamics on signed directed graphs
Douglas Zhou, Shanghai Jiao Tong University, Shanghai, China (10:00-10:30): From Dendrites to Networks: Linking Structure, Dynamics, and Function in Neuronal Systems
Coffee Break (10:30-11:00)
Friedrich Schuessler, TU Berlin, Berlin, Germany (11:00-11:30): Low-rank connectivity structures from learning in recurrent neural networks.
Session: Function and Structure (Chair: Kayson Fakhar)
Petra Vertes, University of Cambridge, Cambridge, UK (11:30-12:00): Reservoir Computing as a Window into Structure-Function Relationships in Neural Systems
Lidia Ripoll-Sánchez, University of Cambridge, Cambridge, UK (12:00-12:30): Molecular connectomes: building the neuropeptide connectome
Speakers' dinner (19:00, individual schedules permitting)
Wednesday, July 9, 2025 (09:00-12:30)
Session (continued): Function and Structure
Anna Levina, University of Tübingen, Tübingen, Germany (9:00-9:30): E/I connectivity shapes outcomes of synaptic plasticity
Session: Dynamics and Function (Chair: Wilhelm Braun)
Katharina Wilmes, University of Bern, Bern, Switzerland (9:30-10:00): Predictive processing in uncertainty
Leo Agueci, ENS, Paris, France (10:00-10:30): A distributed learning framework for motor adaptation
Coffee Break (10:30-11:00)
Agostina Palmigiano, Gatsby Computational Neuroscience Unit, UCL, London, UK (11:00-11:30): Theory of responses to holographic optogenetics in recurrent circuits
Melika Payvand, University of Zürich and ETH Zürich, Zürich, Switzerland (11:30-12:00): Form and Function in Neural Systems: Bidirectional Insights from Neuromorphic Engineering
Panel discussion (12:00-12:30, moderated by Claus C. Hilgetag)
Gorka Zamora-Lopez: Making (some) sense of the structure-function relationship in brain connectivity
For long time, neuroscience emphasized the specialization of different brain regions into particular sensory modalities and tasks. In the last thirty years, a closer study of the white matter has revealed that these specialised regions do not work in isolation, instead, they constantly "talk to each other" via a vast and complex network of communication. This scenario opened the question of how far does the anatomical network determine the neural activity observed (e.g., via functional imaging) is still a matter of debate and speculation. Here, we will shed some light by clarifying that the answer to this "structure-function relation" is not unique or deterministic; for the good and for the bad. On the one hand, the emergent dynamics in a network depend on multiple factors—as it can be shown either from simple propagation dynamics or from more sophisticated whole-brain models—bringing "noise" and methodological limitations to the formulation of answers to the question. On the other hand, the brain constantly engages in different functionalities and transitions across various states; either naturally or as the consequence of dysfunction or damage. This underlines that comprehensive answers to the "structure-function relation" problem shall necessarily accommodate the structural, dynamical and functional flexibility of the brain.
Mariia Popova: Topological building blocks define attractors in a model of excitable brain dynamics on signed directed graphs
Network topology influences the asymptotic dynamical patterns of the network, its attractors, in a general model of excitable dynamics, a three-state susceptible-excited-refractory model, on signed directed graphs. We propose a framework where network topology manifests itself as an interplay of positive and negative feedback loops. A small change in a feedback loop due to an addition or removal of a network edge can drastically change the dynamical patterns of the network, characterized by the appearance and disappearance of attractors from the attractor space of the network. Through a set of numerical experiments, we identify the determinants of such changes in network topology and propose a systematic way to alter the global network dynamics.
As an application example, we discuss the basal ganglia network that is relevant in the context of Parkinson's disease (PD). We show that a given attractor in the attractor space of a network can be induced or destroyed via a specific set of topological manipulations. In the context of PD, this is relevant when talking about perturbation therapies used as a standard method for treatment at the late stages of the disease. As some of the PD symptoms tend to respond to specific stimulation targets better than to others, it is desirable to infer those sites with the help of computational models. For this purpose, a framework that could shape an attractor space of the network and, thus, its dynamical landscape based on the removal of certain edges is of great help. Overall, in this project, we propose a systematic way to alter the dynamical landscape of the system via changes in its topology and, hence, for perturbation therapies like deep brain stimulation in PD.
Douglas Zhou: From Dendrites to Networks: Linking Structure, Dynamics, and Function in Neuronal Systems
Neuronal networks are complex systems where structure, dynamics, and function are deeply intertwined, yet understanding their interplay remains a longstanding challenge. In this talk, I will present a unified perspective spanning multiple scales, from single-neuron dynamics to network-level phenomena. First, I introduce a novel asymptotic framework for modeling neurons with dendritic structures, reducing complex cable models to efficient point neuron models that capture dendritic computation and enable brain-inspired artificial intelligence. Second, I explore causal connectivity in networks, establishing a mapping between structural and causal connectivity using spike trains. By comparing four causality measures, I demonstrate how pairwise analysis can reconstruct structural connectivity, offering a robust methodology for mapping neural circuits. Finally, I investigate balanced states in neuronal networks, where excitatory and inhibitory inputs nearly cancel, driven by strong fluctuations that generate irregular spiking activity. Addressing unresolved questions, I examine heterogeneity in balanced states, and their role in efficient coding and network structure. Together, these insights highlight the importance of multiscale approaches in linking structure, dynamics, and function, offering new directions for understanding both healthy and pathological brain states. This talk underscores how addressing old challenges with innovative tools can drive transformative discoveries in neuroscience.
Friedrich Schuessler: Low-rank connectivity structures from learning in recurrent neural networks.
Recent studies have shown that neural dynamics during experimental tasks is concentrated on low-dimensional manifolds. How such dynamics arise in the brain despite the large number of neurons and potentially large amounts of unstructured connectivity is still not well understood. I will present theoretical work that gives insight into how low-dimensional tasks lead to low-rank components in trained recurrent neural network models. First, the theory helps to understand how these components are related to the task. Second, it shows how the underlying, existing, random connectivity can be exploited by the low-rank connectivity, leading to smaller connectivity changes during learning. Third, I will discuss a simple mechanism which controls the alignment between modes that dominate the network activity (leading principal components) and those that generate the output needed for the task. Taken together, the theory illucidates the functional link between connectivity and dynamics in the brain.
Petra Vertes: Reservoir Computing as a Window into Structure-Function Relationships in Neural Systems
Reservoir computing (RC) provides a powerful framework to probe the computational properties of neural systems. In this talk, I will present two applications of RC in neuroscience, illustrating its utility across multiple scales and experimental paradigms.
First, we apply RC to quantify task performance in a spiking neural network model trained to perform classification and other cognitively relevant tasks. This network incorporates spike-timing-dependent plasticity (STDP), heterogeneous synaptic delays, and exhibits rich temporal dynamics including polychronization. We use RC to relate task performance to the statistics of polychronous groups and to the evolving dynamical state of the network, across diverse configurations involving variations in architecture, delay distributions, excitation-inhibition balance, and neuronal cell types.
In the second part, we extend RC to analyze performance in biologically grounded networks, including large-scale human brain networks derived from diffusion MRI tractography, as well as in-vitro neuronal cultures from murine cortex and human cerebral organoids. In each case, the physical or inferred connectivity structure is used as a reservoir, and we assess the system’s intrinsic ability to support task-relevant computations.
Together, these studies demonstrate how RC can serve as a unified framework to explore structure-function relationships in neural circuits, bridging the gap between artificial and biological computation and offering insights into neural coding and mechanisms of dysfunction in disease.
Lidia Ripoll-Sánchez: Molecular connectomes: building the neuropeptide connectome
Neuropeptides modulate synaptically wired neuronal circuits, playing a critical role in nervous systemfunction. Yet, little is known about the structure and function of extrasynaptic signalling networks at awhole-organism level, or how these networks are maintained across evolution.
To this end, we used single neuron gene expression and deorphanisation data from neuropeptide-activated G-protein coupled receptors to generate a connectome of 92 neuropeptide signalling networks in several Caenorhabditis species at different developmental stages. These networks defined a connection when the sending neuron expressed a neuropeptide, the receiving neuron expressed the cognate receptor, and both neurons extended overlapping processes. We then applied graph theory and machine learning methods to characterise the structural features of both the individual networks and the integrated network formed by combining them.
Our analysis on the connectivity pattern revealed a mesoscale structure for the core of this network, splitting it in three groups of neurons that act as highly controlled functional hubs. Notably whithin these hubs, we identified a group of neurons that seem to be morphologically and biochemically adapted for neuropeptidergic communication. Comparing across developmental stages and Caenorhabditis species, we found that this structured network emerges early in development and shows strong conservation across species. Moreover, the neuropeptide network exhibits a higher connection density than synaptic and gap junction networks, connecting neurons that are not synaptically connected.
These findings challenge the idea that neuronal communication is primarily synaptic, revealing a dense, decentralised neuropeptide network with functional and structural roles. Conserved signalling patterns across species further highlight the evolutionary significance of neuropeptide connectivity. We expect that this newly mapped neuropeptide connectomes, their analysis and the interactive website we developed to explore them (nemamod.org) will serve as a prototype for other animals, providing new insights into the structure of neuromodulatory networks in larger brains.
Katharina Wilmes: Predictive processing in uncertainty
Prediction error neurons have been identified in layer 2/3 of diverse brain areas. In the first part of this talk, I will address how uncertainty could modulate prediction error activity and hence learning. Here, we use a normative approach to derive how prediction errors should be modulated by uncertainty and postulate that such uncertainty-modulated prediction errors (UPE) are represented by layer 2/3 pyramidal neurons. By implementing the calculation of UPEs in a microcircuit model, we show that different cell types in cortical circuits can compute means, variances and UPEs with local activity-dependent plasticity rules. Finally, we show that the resulting UPEs enable adaptive learning rates.
Leo Agueci : A distributed learning framework for motor adaptation
Adaptation is a fundamental aspect of motor learning. Intelligent systems must adapt to perturbations in the environment while simultaneously maintaining stable memories. Classic work has argued that this trade-off could be resolved by complementary learning systems operating at different speeds; yet the mechanisms enabling coordination between slow and fast systems remain unknown. Here, we propose a multi-region distributed learning model in which learning is shared between two populations of neurons with distinct roles and structures: a recurrent "controller" network which stores a slowly evolving memory, and a feedforward "adapter" network that rapidly learns to respond to perturbations in the environment. In our model, supervised learning in the adapter produces a predictive error signal that simultaneously tutors consolidation in the controller through a local plasticity rule. Our model offers insight into the mechanisms that may support distributed computations in the motor cortex and cerebellum during motor adaptation
Agostina Palmigiano: Theory of responses to holographic optogenetics in recurrent circuits
The cerebral cortex exhibits strong recurrent connectivity governed by complex wiring rules. Recent advances allow probing this circuitry through optogenetic perturbations of single or few neurons, yet a general theory linking these responses to underlying connectivity is lacking. Here, we develop an exactly solvable framework to predict responses to perturbations in networks with multiple cell types and space- and feature-dependent connectivity. Analysis of these equations reveals simple rules that govern perturbation response patterns. Comparison with experimental data imposes new constraints on the connectivity parameters and enables predictions of the responses of unmeasured cell types and novel experimental conditions.
Melika Payvand: Form and Function in Neural Systems: Bidirectional Insights from Neuromorphic Engineering
The correlation between form and function in neural systems can be thought of in two ways.
First, by drawing inspiration from the observed morphology and architecture of biological brains, we can incorporate these features as inductive biases and explore their role in computation.
In this domain, we, as neuromorphic engineers, explore hardware plausible neural circuits and architectures which when used as an inductive bias bring efficiency and performance to our systems while solving a real-world task. Specifically, I will talk about our exploration of these questions in the case of morphology of dendrites and specific local connectivity of neurons.
Second, we can approach the problem in the reverse direction- by imposing the desired functionality on the system, applying constraints akin to those in neural systems (such as limited memory and energy) and observe whether the optimization process finds a meaningful structure. This is of course a much more ill-defined and open-ended problem. I will talk about some interesting observation we have found in this domain, though largely inconclusive.
Anna Levina: E/I connectivity shapes outcomes of synaptic plasticity
Typical studies of learning in biological neural networks focus on how plausible plasticity rules shape individual synaptic connections, either in recurrent networks or feed-forward projections. However, in the latter case, presynaptic network properties—such as clustered connectivity—strongly influence population dynamics and, consequently, the learning. Here, we address this gap by demonstrating how non-plastic network structures can significantly impact synaptic learning outcomes. We show that unstructured recurrent connectivity and noise can impair the ability of synaptic plasticity to separate presynaptic input populations and to coordinate excitatory and inhibitory connection development. In contrast, introducing overlapping clustered structures in fixed or plastic recurrent connectivity enhances synaptic learning. Using Bayesian inference, we identify optimal recurrent network connectivity patterns that maximize learning efficiency and demonstrate how relatively simple connectivity structures can profoundly shape network plasticity. As a next step, we investigate how the coordination of excitatory and inhibitory inputs on the dendritic tree, along with the shape of synaptic nonlinearities, influences the emergence of tuning under synaptic plasticity.
