Symposium
Data-Driven Approaches to Understanding Mechanisms of Brain Function
Chair
Emiliano Marachlian
Departamento de Fisica ,FCEyN,UBA - IFIByNE, UBA-CONICET
Co-Chair
Mariana Luz Bavassi
IFIByNE, UBA-CONICET - Departamento de Fisica ,FCEyN,UBA
Access to increasingly large and high-quality datasets is rapidly transforming neuroscience. Modern techniques for manipulating and measuring gene expression, imaging, optogenetics, electrophysiology, as well as the application of artificial intelligence to data analysis, are providing unprecedented opportunities to study how the brain works and to address questions that were previously inaccessible. However, the growth in both the quantity and quality of available information also demands that scientists develop new theoretical frameworks and technical tools capable of integrating and interpreting these data, redirecting questions and experiments, and allowing us to fully exploit their potential.
In this symposium we will present four different perspectives, ranging from modeling, optogenetics, and machine learning to the analysis of behavioral experiments. Together, these contributions aim to provide a broad overview of the strategies and ideas that are shaping modern neuroscience. We hope this discussion will also help open a broader reflection on what the role of Argentine neuroscience can and should be in this ongoing scientific transformation.
Agostina Palmigiano
Gatsby Computational Neuroscience Unit, Londres, UK
Theories of responses to optogenetic perturbations.
Resumen Charla Disertante 1 (en inglés): Optogenetics has become a powerful tool to causally probe neural circuits by precisely controlling neuronal activity. In this talk, I will discuss how combining optogenetic perturbations with theoretical and data-driven approaches can help uncover fundamental principles of brain function. Rather than focusing on specific experimental details, I will highlight how responses to controlled perturbations provide insight into the structure and dynamics of neural networks. By interpreting these responses through the lens of theoretical models, we can begin to infer how collective circuit properties shape activity patterns and constrain neural computations. This approach illustrates how modern neuroscience increasingly relies on the interplay between large-scale data, targeted manipulations, and theory to move from observation to mechanistic understanding.
Claudio Mirasso
Institute for Cross-Disciplinary Physics and Complex Systems (IFISC -UIB-CSIC), Palma de Mallorca, España.
Data-Driven Biorealistic Models of Mouse Visual Cortex Dynamics
I’ll present a biologically realistic model of the mouse visual cortex comprising both primary visual cortex (V1) and the Lateromedial (LM) visual areas. The model combines datasets on cell-type composition, intrinsic electrophysiology, and connectomics, and uses deep learning to optimize synaptic weights so as to reproduce in vivo Neuropixels recordings. This provides a data-driven, biologically grounded description of circuit dynamics. We will use it to investigate the emergent properties of the coupled V1–LM system and quantify the functional impact of feedback from LM to V1. By comparing isolated V1 models with the full V1–LM circuit, we will highlight differences in receptive field size, time to first spike, and pattern recognition performance. These findings shed light on how top-down feedback shapes early sensory processing and supports computation in biologically realistic cortical networks.
Cecilia Martinez
Laboratorio de Fisiología y Algoritmos del Cerebro, Fundación Instituto Leloir, IBBA-CONICET
Machine Learning Approaches for Studying Learning and Memory
Understanding how animals learn and remember requires a precise account of how their actions change over time. However, behavior has historically been treated as a "black box," measured through coarse, aggregate metrics like trial duration or time spent in a zone. These simplified snapshots often mask the complex, sub-second behavioral dynamics that characterize the actual process of learning.
Recent advances in machine learning are revolutionizing behavioral quantification by allowing us to treat movement as a continuous, high-dimensional data stream. In this talk, I will discuss how markerless pose estimation (e.g., DeepLabCut) and unsupervised behavioral segmentation can be used to "deconstruct" complex tasks. By extracting precise body postures and identifying hidden behavioral states, we can observe the fine-grained evolution of strategies during spatial exploration and object-based recognition. These ML-driven approaches reveal that learning is not just a change in an outcome metric, but a fundamental reorganization of behavioral architecture.
Joao Barbosa
Neuromodulation Institute and Neurospin, París, Francia.
Region-specific transformations enable distributed computations of flexible decisions.
Cognitive flexibility is thought to rely on the prefrontal cortex controlling other regions. However, recent evidence shows that information required to compute flexible decisions is distributed across multiple regions (“everything is everywhere”, 2025; International Brain Laboratory, 2025). This shift raises some questions: is ’everything is everywhere’ compatible with different task variables being first computed within specific regions and then broadcasted? If so, how to identify which information is communicated across specific regions to ultimately enable flexible decisions?
Here, we tackle these questions by analyzing the dynamics within and between six brain regions of the monkey brain engaged in a context-dependent decision-making task (Mante & Susillo et al., 2013). Decoding from this dataset has previously found task-relevant information across all the recorded regions (Siegel et al., 2015). Using population analyses and data-constrained recurrent neural networks (RNNs) we found significant differences in within-region dynamics not captured by classical decoding analyses.
Inspecting communication subspaces (Semedo et al., 2019) inferred with low-rank RNNs, we found that the dimensionality of within regional dynamics was higher than what is communicated downstream in sensory regions (e.g. V4 and MT), replicating previous results in early visual cortex (Semedo et al., 2019).
In contrast, within- and across-regional dimensionality did not differ in frontal regions (i.e. LIP, FEF and PFC). Additionally, we found that overall encoding dimensionality reduced with hierarchical position, suggesting progressively more abstract transformations as information propagated downstream. Finally, in-silico perturbations revealed how choice computation abruptly emerged with stronger cross-regional interactions, in contrast to stimuli encoding which remained stable. Perturbing inputs to sensory regions confirmed a bottom-up flow of task-stimuli into the frontal regions. Altogether, using population analyses that go beyond decoding and data-constrained RNN we show that despite task-variables being broadly encoded, each region performs different computations during flexible decision making.