Neural manifolds

Juan Gallego discusses whether neural manifolds are real explanatory structures and how population activity relates to behaviour. Topics include causal interpretations, evolutionary claims, motor control and the limits of low-dimensional descriptions, along with work on restoring movement after spinal-cord injury.

…among criticality, sleep and neural manifolds, drawing on their review of the experimental literature.

…place cells, forming low-dimensional neural manifolds that support flexible navigation. However, beyond simple line and ring attractors, how the brain represents and navigates more complex neural manifolds remains unclear. In this study, we propose a solution for representing arbitrary nonlinear neural manifolds by approximating them with piecewise linear spaces composed of simplexes, known as “simplicial compl

Recent studies have proposed that neural circuits generate a task manifold: i.e., a subset of the neural state-space to which neural activity is confined as an animal performs a task. Thus, discovering and characterising these manifolds from experimental data can shed light on the neural computations unfolding within the brain during behavioural tasks. In parallel, neuroscience has seen a recent surge of interest into fitting dynamical systems to experimental data with the goal of reverse-engineering neural circuits. Yet, common neural manifolds frameworks often do not take into account that n

In the brain, the collective activity of neurons form trajectories on low dimensional neural manifolds. The neural computation underlying flexible cognition and behavior relies on dynamic control of these structures. For example, different tasks or behaviors are represented on different subspaces, requiring fast timescale subspace rotation to move from one behavior to the next. For flexibility in a particular behavior, the neural trajectory must be dynamically controllable within that behaviorally determined subspace. Here we propose mechanisms for fast timescale control of the neural subspace

Brain activity depends on ongoing behavioural and arousal states, such as attention, eye closure, or sleep. These states display measurable differences in population activity [1], dimensionality [2], noise correlations [3], and autocorrelation timescales [4]. State-dependent changes in the primary visual cortex (V1) are potentially induced by cortico-cortical signals, such as those from V4 to V1, known to mediate visual attention in macaques [5]. However, relationships between such top-down signals and the aforementioned state-dependent changes have not been demonstrated directly. To investig

A major challenge in studying spatial memory is understanding whether and how different navigational strategies influence hippocampal spatial representations of the environment.We designed a spatial recency memory task with two variants: one requiring allocentric representations of location, the other entailing egocentric representations. 10 freely-moving rats were trained over 34 daily sessions; each day, rats learned and then recalled the position of a food reward among six locations. Each trial, animals entered the arena from one of four start locations, generating distinct origin-goal traj

Computations in the brain and artificial neural networks can be understood as dynamical processes conformed by the activity of large neural populations. Yet revealing the structure of the underpinning latent dynamical processes from data and interpreting their relevance in computational tasks remains a fundamental challenge. Recent evidence suggests that task-relevant neural activity takes place on low-dimensional subspaces of the state space called neural manifolds. However, there is a lack of theoretical frameworks for the unsupervised representation of neural dynamics that are interpretable

…with recent evidence for “neural manifolds”: low-dimensional, functionally-relevant descriptions of population activity. The most general way to discover manifolds is to find the simplest transform of activity that maximises the mutual information (MI) between the transformation and the relevant task variables, since MI is the most general meas

…and the widespread adoption of neural population frameworks, there is an increased need to achieve efficient decoding of task-relevant behavioral information from neural dynamics. The most prevalent methods to uncover task-related topologies in population dynamics, often approximated as neural manifolds, are linear dimensionality reduction methods such as principal component analysis (PCA). These techniques, however, find large-variance subspaces within which to embed neural manifolds and such embeddings do not necessaril

…the ‘black box’ of representations in neural circuits. In this talk, we will demonstrate theoretical approaches that help describe how cognitive and behavioral task implementations emerge from the structure in neural populations and from biologically plausible neural networks. First, we will introduce an analytic theory that connects geometric structures th

The activity of neural populations can be well-described by relatively few population-wide activity patterns spanning a “neural manifold”. Virtually all these studies have analysed flat neural manifolds to understand how the brain controls behaviour. We hypothesised that since neurons have nonlinear responses and make thousands of recurrent connections that may enhance this nonlinearity, nonlinear manifolds should capture the neural population activity better than flat manifolds. Analysis of a centre-out reaching task in monkeys confirmed that, even during a relatively simple behaviour, moto

While neural population recordings are increasingly high-dimensional, the bulk of recorded activities is usually structured along lower-dimensional manifolds. These manifolds often appear to be nonlinear, so that linear dimensionality reduction methods such as principal component analysis (PCA) yield embedding dimensionalities (ED) that are much higher than the intrinsic dimensionality (ID) of the manifold [Jazayeri, Ostojic, 2021]. Across datasets, the embedding often generates components that resemble higher-order functions of each other, or ‘higher-order components’ (HOCs), thus suggesting

Neural circuit computations are modulated by combinations of internal and external factors. For example, neurons in the medial entorhinal cortex (MEC) change their firing patterns or “remap” during a reward-seeking task, responding to changing task conditions and environmental cues. Why does this remapping occur? What unifying principles (if any) underlie remapping? Recent work proposes that remapping reflects hidden state inference—changes to an animal’s internal beliefs about the world. Recent experimental evidence suggests that remapping can indeed arise from latent behavioral or cognitive

…emerge from structure in neural populations and from biologically plausible learning rules. First, I will introduce an analytic theory that connects geometric structures that arise from neural res

…a new class of multi-region recurrent neural network (RNN) models to infer brain-wide interactions driving such maladaptive behaviour. The models were trained to match experimental data across two levels simultaneously: brain-wide neural dynamics from 10-40,000 neurons and the realtime behaviour of the fish. Analysis of the trained R