Symposium Program

Further details on the symposium program will be available soon.

Time	Session
9:30 – 9:45	Introduction
9:45 – 10:20	Speaker 1
10:20 – 10:55	Speaker 2
10:55 – 11:25	Coffee Break
11:25 – 12:00	Speaker 3
12:00 – 12:35	Speaker 4
12:35 – 14:00	Lunch + Poster Session
14:00 – 14:35	Speaker 5
14:35 – 15:10	Speaker 6
15:10 – 15:45	Speaker 7
15:45 – 16:15	Coffee Break
16:15 – 16:50	Speaker 8
16:50 – 17:25	Speaker 9
17:25 – 17:40	Concluding Remarks

List of confirmed speakers (alphabetical order):

1. Lorenzo Gaetano Amato, Sant'Anna School of Advanced Studies, Pisa, IT
2. Demian Battaglia, University of Strasbourg, FR
3. Daniel Durstewitz, Central Institute for Mental Health, Mannheim, DE
4. Tatiana A. Engel, Princeton University, Princeton, USA
5. Sean Froudist-Walsh, University of Bristol, Bristol, UK
6. Matthieu Gilson, Institut des Neurosciences, Marseille, FR
7. Maurizio Mattia, Istituto Superiore di Sanità, Roma, IT
8. Leonardo Pollina, École Polytechnique Fédérale de Lausanne, Lausanne, CH
9. Sara Solla, Northwestern University, Evanston, USA

Abstracts

Digital Twins Enable Early Disease Diagnosis by Reconstructing Neurodegeneration Levels from Neural Recordings Lorenzo Gaetano Amato, Sant'Anna School of Advanced Studies, Pisa, IT Understanding how structural alterations in the brain lead to functional anomalies is a central challenge in neuroscience and clinical neurology. This talk will discuss an emerging approach, leveraging computational modeling to derive personalized digital biomarkers from non-invasive neural recordings. These biomarkers estimate individual brain pathology by inverting individual experimental data, such as EEG recordings. Specifically, we use a personalized whole-brain modeling framework to estimate biophysically meaningful parameters that capture the progression of neurodegenerative alterations, translating non-invasive electrophysiological data into mechanistic insights about disease. We demonstrate this approach in the context of Alzheimer’s disease, where we apply the Digital Alzheimer’s Disease Diagnosis (DADD) model to a large cohort of EEG recordings. The derived digital biomarkers successfully reflect underlying pathological changes, predict positivity to biological markers, and significantly improve the diagnostic and prognostic power of EEG. This model-based strategy opens new possibilities for affordable, scalable, and non-invasive monitoring of brain health, potentially enabling earlier detection and stratification of patients along the disease continuum.

Toward Foundation Models for Dynamical Systems Reconstruction in Neuroscience Daniel Durstewitz, Central Istitute for Menthal Health, Germany Rather than hand-crafting computational theories of neural function, recent progress in scientific machine learning (ML) and AI suggests that we may be able to infer dynamical-computational models directly from neurophysiological and behavioral data. This is called dynamical systems reconstruction (DSR), the learning of generative surrogate models of the underlying dynamics from time series observations. In my talk I will cover recent ML/AI architectures, training algorithms, and validation procedures for DSR, and how they can integrate neuroscience data from multiple modalities, animals, and task designs, into a joint latent model. Finally, I will introduce DynaMix, a recent interpretable DSR foundation model which exhibits the zero-shot inference and in-context learning capabilities known from LLMs.

The dynamics and geometry of choice in the premotor cortex Tatiana A. Engel, Princeton University, Princeton, USA Neural responses in association brain areas during cognitive tasks are heterogeneous, and the widespread assumption is that this heterogeneity reflects complex dynamics involved in cognition. However, the complexity may arise from a fundamentally different coding principle: the collective dynamics of a neural population encode simple cognitive variables, while individual neurons have diverse tuning to the cognitive variable, similar to tuning curves of sensory neurons to external stimuli. We developed an approach to simultaneously infer neural population dynamics and tuning functions of single neurons to the latent population state. Applied to spike data recorded from primate premotor cortex during decision-making, our model revealed that populations of neurons encoded the same dynamic variable predicting choices, and heterogeneous firing rates resulted from the diverse tuning of single neurons to this decision variable. The inferred dynamics indicated an attractor mechanism for decision computation. Our results reveal a unifying geometric principle for neural encoding of sensory and dynamic cognitive variables.

Associative RNNs with self-interactions: reservoir computing without catastrophic forgetting Maurizio Mattia, Istituto Superiore di Sanità, Roma, IT Associative memory models form a key link between statistical physics and theoretical neuroscience. Traditionally, spin-glass models with Hebbian learning in absence of self-interactions suffer catastrophic forgetting when memory load exceeds a critical threshold. Here I will show that including Hebbian self-couplings in deterministic, graded-unit recurrent neural networks (RNNs) fundamentally reshapes the energy landscape, confining dynamics to the low-dimensional space of stored patterns. This enables robust recall at any memory load, overcoming catastrophic forgetting without altering the Hebbian matrix or requiring nonlocal learning. Beyond storing static patterns, these RNNs can implement reservoir computing to learn and recall complex dynamical sequences, even if the starting Amari-Hopfield synaptic matrix is symmetric. Our theory elucidates how such networks mimic arbitrary dynamical systems evolving within low-dimensional state spaces and predicts optimal parameters for performance. In conclusion I will demonstrate how these RNNs effectively replicate premotor cortical activity recorded from monkeys performing a stop-signal task, offering novel insights into motor decision and movement inhibition.

Disentangling neural representations underlying movement execution and imagery for human augmentation Leonardo Pollina, École Polytechnique Fédérale de Lausanne, Lausanne, CH One of the central challenges in human motor augmentation is the neural resource allocation problem: how to identify reliable control signals in physiological activity that could drive external devices without interfering with ongoing biological functions. Neural population dynamics and latent neural representations offer a promising avenue for addressing this challenge by allowing us to disentangle overlapping motor processes. In this talk, I will explore how motor execution and motor imagery are represented in electrocorticography (ECoG) data recorded from a tetraplegic patient with partial upper-limb residual function. I will then introduce an approach that involves identifying distinct neural subspaces, specific to either executed or imagined movements, as a step toward isolating viable control signals and advancing the development of intuitive brain-machine interfaces for both motor restoration and augmentation.