A unifying dynamical theory of distributed computation and generalisation in biological and artificial neural systems

Descripción del proyecto

Descodificación de los cálculos encefálicos

Se sabe que el encéfalo realiza cálculos complejos, pero nuestra comprensión de cómo las redes neuronales distribuidas trabajan juntas para lograrlo es limitada. El objetivo fundamental del proyecto NEURO-FUSE, financiado por el Consejo Europeo de Investigación, es delinear la contribución de neuronas individuales a procesos cognitivos más amplios y distribuidos por varias regiones encefálicas. La hipótesis de trabajo es que la dinámica encefálica general entra en distintos estados atractores durante las tareas, lo que crea representaciones invariantes independientes de neuronas específicas. Al inferir estas invariantes, el proyecto reconstruirá la actividad de todo el encéfalo a partir de grabaciones locales en ratones y macacos. El aprendizaje automático ayudará a revelar procesos cognitivos distribuidos, lo que superará las limitaciones del muestreo local en neurociencia.

Objetivo

Understanding how the coordinated activity of neurons in multiple brain regions achieves robust behaviour is one of the most fundamental questions in neuroscience. Although recent single-cell technologies enable addressing this question by recording from large neural populations, they are limited to surveying focal brain regions and superficial cortical layers. Without an analytical framework to jointly model isolated measurements, we cannot hope to understand and quantitatively model how single-neuron dynamics give rise to distributed computations. I hypothesise that global brain dynamics fall on distinct attractor states during a given stimulus or task. Attractors naturally give rise to invariant representations, dynamical motifs independent of the sampled neurons’ identity. Inferring these invariances would allow reconstructing activity in extended regions from incomplete local recordings to reveal brain-wide cognitive processes. Further, composing invariances would provide insights into the neural correlates of generalisation, with a broad impact on neuroscience and machine learning. I propose a novel mathematical theory combining abstract combinatorial dynamical systems theory and modern machine learning to infer and compose invariant latent dynamics across measurements. We will use this theory to unify large-scale cell-resolution recordings of the mouse and macaque cortex into a common model to make cell-specific predictions across several brain regions. Our results could fundamentally challenge our view on distributed cognitive computations by revealing moment-by-moment single-neuron dynamics in spatially distributed neurons. Further, my theory will help understand how the brain generalises knowledge across tasks by composing and repurposing invariances. More broadly, my theory will open new avenues for machine learning and neuroscience to interact through sharing and shaping the dynamical processes that underpin neural computations in vivo and in silico.