Daniel Durstewitz (Zentral Institut für seelische Gesundheit & Interdisciplinary Center for Scientific Computing at Heidelberg University):

Dynamical Systems Foundation Models for Neuroscience

Abstract: 

Foundation models like, e.g., large language models, are models that are pre-trained on a large corpus of data with diverse applications. They accomplish feats like in-context learning (learning to extrapolate just from examples without parameter fine-tuning) and zero-shot generalization to novel domains and tasks.
Their promise in neuroscience lies with i) facilitating the discovery of common computational motifs and principles across diverse tasks and datasets, and ii) providing zero-shot predictions for novel experimental conditions. For these purposes, and – more generally – to be useful as neuroscientific discovery tools that can provide insight into computational mechanisms, they need to be interpretable and, ideally, mathematically tractable. 
In my talk I will present two types of foundation model architectures based on dynamical systems principles that fulfill these criteria, and that are able to few- or zero-shot reconstruct unseen dynamical systems just from “in-context data”.

Rosanne Rademaker (Ernst Strüngmann Institute for Neuroscience):

Visual working memory, and how its representations best serve human behavior 

Abstract: 

Neural correlates of visual working memory have been found in early visual, parietal, and prefrontal regions. These findings have spurred fruitful debate over how and where in the brain memories might be represented. Here, I will present data from multiple experiments to demonstrate how a focus on behavioral requirements can unveil a more comprehensive understanding of the visual working memory system. Specifically, items in working memory must be maintained in a highly robust manner, resilient to interference. At the same time, storage mechanisms must preserve a high degree of flexibility to meet changes in behavioral goals. Several examples will be explored in which visual memory representations are shown to undergo transformations, and even shift their cortical locus alongside their coding format based on specifics of the task.

Jean Laurens (Ernst Strüngmann Institute for Neuroscience): 

Head Direction Cells: the 3D Neuronal Compass - and a Tour through Laurens Lab   

Abstract: 

The head direction cell (HDC) system is a neuronal circuit that encodes an animal’s orientation in the horizontal plane, functioning as an internal neural compass. Although traditionally studied in planar environments, recent findings show that this system can also represent head orientation in three dimensions. In this talk, I will outline the geometry of 3D head rotations and introduce theoretical models for 3D orientation encoding. I will then present a series of experiments in mice demonstrating robust 3D head-orientation coding within the HDC circuit, and examine their underlying sensory mechanisms as well as the extent to which these findings generalize across mammalian species. 

I will conclude by presenting my laboratory at the Ernst Strüngmann Institute in Frankfurt, where I carry related work in rats and marmoset monkeys. 

13:00 GSC 0.06/0.07

Neural population dynamics are shaped by many cellular, synaptic, and network properties. Not only is it important to understand how coordinated changes in circuit parameters alter neural activity, but also when dynamics are unaffected by—or invariant to—such changes. Computational modeling has revealed invariances in single neurons and small circuits that are thought to reflect their robustness against variability and perturbations. However, generalizing these insights to larger circuits in cortex and other brain areas remains challenging. 

A key bottleneck lies in inverse modeling of neural circuits with spiking network models, i.e., identifying parameter configurations that quantitatively match dynamics observed in neural recordings. Here, we present Automated Model Inference from Neural Dynamics (AutoMIND) for efficient discovery of invariant circuit model configurations. AutoMIND leverages a mechanistic model with adaptive spiking neurons and clustered connectivity, which displays a rich variety of spatiotemporal dynamics. Probabilistic deep generative models—trained on network simulations only—then returns many parameter configurations consistent with a given target observation of neural activity. 
Applied to several datasets, AutoMIND discovers circuit models of synchronous network bursting in human brain organoids across early development, as well as models capturing complex frequency profiles of Neuropixels recordings in mouse hippocampus and cortex. In each case, we obtain hundreds of configurations that compose a (nonlinear) parameter subspace in which population dynamics remain unchanged. Surprisingly, global and local geometries of the invariant subspace are not fixed, but differ for different dynamics. Together, our results shed light on dynamic-dependent invariances of circuit parameters underlying diverse population dynamics, while demonstrating the flexibility of AutoMIND for inverse modeling of neural circuits.