Representations and generalization in artificial and brain neural networks.

Humans and animals excel at generalizing from limited data, a capability yet to be fully replicated in artificial intelligence. This perspective investigates generalization in biological and artificial deep neural networks (DNNs), in both in-distribution and out-of-distribution contexts. We introduce two hypotheses: First, the geometric properties of the neural manifolds associated with discrete cognitive entities, such as objects, words, and concepts, are powerful order parameters. They link the neural substrate to the generalization capabilities and provide a unified methodology bridging gaps between neuroscience, machine learning, and cognitive science. We overview recent progress in studying the geometry of neural manifolds, particularly in visual object recognition, and discuss theories connecting manifold dimension and radius to generalization capacity. Second, we suggest that the theory of learning in wide DNNs, especially in the thermodynamic limit, provides mechanistic insights into the learning processes generating desired neural representational geometries and generalization. This includes the role of weight norm regularization, network architecture, and hyper-parameters. We will explore recent advances in this theory and ongoing challenges. We also discuss the dynamics of learning and its relevance to the issue of representational drift in the brain.

Common population codes produce extremely nonlinear neural manifolds.

Populations of neurons represent sensory, motor, and cognitive variables via patterns of activity distributed across the population. The size of the population used to encode a variable is typically much greater than the dimension of the variable itself, and thus, the corresponding neural population activity occupies lower-dimensional subsets of the full set of possible activity states. Given population activity data with such lower-dimensional structure, a fundamental question asks how close the low-dimensional data lie to a linear subspace. The linearity or nonlinearity of the low-dimensional structure reflects important computational features of the encoding, such as robustness and generalizability. Moreover, identifying such linear structure underlies common data analysis methods such as Principal Component Analysis (PCA). Here, we show that for data drawn from many common population codes the resulting point clouds and manifolds are exceedingly nonlinear, with the dimension of the best-fitting linear subspace growing at least exponentially with the true dimension of the data. Consequently, linear methods like PCA fail dramatically at identifying the true underlying structure, even in the limit of arbitrarily many data points and no noise.

The sloppy relationship between neural circuit structure and function.

Investigating and describing the relationships between the structure of a circuit and its function has a long tradition in neuroscience. Since neural circuits acquire their structure through sophisticated developmental programmes, and memories and experiences are maintained through synaptic modification, it is to be expected that structure is closely linked to function. Recent findings challenge this hypothesis from three different angles: function does not strongly constrain circuit parameters, many parameters in neural circuits are irrelevant and contribute little to function, and circuit parameters are unstable and subject to constant random drift. At the same time, however, recent work also showed that dynamics in neural circuit activity that is related to function are robust over time and across individuals. Here this apparent contradiction is addressed by considering the properties of neural manifolds that restrict circuit activity to functionally relevant subspaces, and it will be suggested that degenerate, anisotropic and unstable parameter spaces are closely related to the structure and implementation of functionally relevant neural manifolds.

Neural manifold analysis of brain circuit dynamics in health and disease.

Recent developments in experimental neuroscience make it possible to simultaneously record the activity of thousands of neurons. However, the development of analysis approaches for such large-scale neural recordings have been slower than those applicable to single-cell experiments. One approach that has gained recent popularity is neural manifold learning. This approach takes advantage of the fact that often, even though neural datasets may be very high dimensional, the dynamics of neural activity tends to traverse a much lower-dimensional space. The topological structures formed by these low-dimensional neural subspaces are referred to as "neural manifolds", and may potentially provide insight linking neural circuit dynamics with cognitive function and behavioral performance. In this paper we review a number of linear and non-linear approaches to neural manifold learning, including principal component analysis (PCA), multi-dimensional scaling (MDS), Isomap, locally linear embedding (LLE), Laplacian eigenmaps (LEM), t-SNE, and uniform manifold approximation and projection (UMAP). We outline these methods under a common mathematical nomenclature, and compare their advantages and disadvantages with respect to their use for neural data analysis. We apply them to a number of datasets from published literature, comparing the manifolds that result from their application to hippocampal place cells, motor cortical neurons during a reaching task, and prefrontal cortical neurons during a multi-behavior task. We find that in many circumstances linear algorithms produce similar results to non-linear methods, although in particular cases where the behavioral complexity is greater, non-linear methods tend to find lower-dimensional manifolds, at the possible expense of interpretability. We demonstrate that these methods are applicable to the study of neurological disorders through simulation of a mouse model of Alzheimer's Disease, and speculate that neural manifold analysis may help us to understand the circuit-level consequences of molecular and cellular neuropathology.

Geodesic-based distance reveals nonlinear topological features in neural activity from mouse visual cortex.

An increasingly popular approach to the analysis of neural data is to treat activity patterns as being constrained to and sampled from a manifold, which can be characterized by its topology. The persistent homology method identifies the type and number of holes in the manifold, thereby yielding functional information about the coding and dynamic properties of the underlying neural network. In this work, we give examples of highly nonlinear manifolds in which the persistent homology algorithm fails when it uses the Euclidean distance because it does not always yield a good approximation to the true distance distribution of a point cloud sampled from a manifold. To deal with this issue, we instead estimate the geodesic distance which is a better approximation of the true distance distribution and can therefore be used to successfully identify highly nonlinear features with persistent homology. To document the utility of the method, we utilize a toy model comprised of a circular manifold, built from orthogonal sinusoidal coordinate functions and show how the choice of metric determines the performance of the persistent homology algorithm. Furthermore, we explore the robustness of the method across different manifold properties, like the number of samples, curvature and amount of added noise. We point out strategies for interpreting its results as well as some possible pitfalls of its application. Subsequently, we apply this analysis to neural data coming from the Visual Coding-Neuropixels dataset recorded at the Allen Institute in mouse visual cortex in response to stimulation with drifting gratings. We find that different manifolds with a non-trivial topology can be seen across regions and stimulus properties. Finally, we interpret how these changes in manifold topology along with stimulus parameters and cortical region inform how the brain performs visual computation.

Emergence of an Invariant Representation of Texture in Primate Somatosensory Cortex.

A major function of sensory processing is to achieve neural representations of objects that are stable across changes in context and perspective. Small changes in exploratory behavior can lead to large changes in signals at the sensory periphery, thus resulting in ambiguous neural representations of objects. Overcoming this ambiguity is a hallmark of human object recognition across sensory modalities. Here, we investigate how the perception of tactile texture remains stable across exploratory movements of the hand, including changes in scanning speed, despite the concomitant changes in afferent responses. To this end, we scanned a wide range of everyday textures across the fingertips of rhesus macaques at multiple speeds and recorded the responses evoked in tactile nerve fibers and somatosensory cortical neurons (from Brodmann areas 3b, 1, and 2). We found that individual cortical neurons exhibit a wider range of speed-sensitivities than do nerve fibers. The resulting representations of speed and texture in cortex are more independent than are their counterparts in the nerve and account for speed-invariant perception of texture. We demonstrate that this separation of speed and texture information is a natural consequence of previously described cortical computations.

Simultaneous, cortex-wide dynamics of up to 1 million neurons reveal unbounded scaling of dimensionality with neuron number.

The brain's remarkable properties arise from the collective activity of millions of neurons. Widespread application of dimensionality reduction to multi-neuron recordings implies that neural dynamics can be approximated by low-dimensional "latent" signals reflecting neural computations. However, can such low-dimensional representations truly explain the vast range of brain activity, and if not, what is the appropriate resolution and scale of recording to capture them? Imaging neural activity at cellular resolution and near-simultaneously across the mouse cortex, we demonstrate an unbounded scaling of dimensionality with neuron number in populations up to 1 million neurons. Although half of the neural variance is contained within sixteen dimensions correlated with behavior, our discovered scaling of dimensionality corresponds to an ever-increasing number of neuronal ensembles without immediate behavioral or sensory correlates. The activity patterns underlying these higher dimensions are fine grained and cortex wide, highlighting that large-scale, cellular-resolution recording is required to uncover the full substrates of neuronal computations.

Multiscale model of primary motor cortex circuits predicts in vivo cell-type-specific, behavioral state-dependent dynamics.

Understanding cortical function requires studying multiple scales: molecular, cellular, circuit, and behavioral. We develop a multiscale, biophysically detailed model of mouse primary motor cortex (M1) with over 10,000 neurons and 30 million synapses. Neuron types, densities, spatial distributions, morphologies, biophysics, connectivity, and dendritic synapse locations are constrained by experimental data. The model includes long-range inputs from seven thalamic and cortical regions and noradrenergic inputs. Connectivity depends on cell class and cortical depth at sublaminar resolution. The model accurately predicts in vivo layer- and cell-type-specific responses (firing rates and LFP) associated with behavioral states (quiet wakefulness and movement) and experimental manipulations (noradrenaline receptor blockade and thalamus inactivation). We generate mechanistic hypotheses underlying the observed activity and analyzed low-dimensional population latent dynamics. This quantitative theoretical framework can be used to integrate and interpret M1 experimental data and sheds light on the cell-type-specific multiscale dynamics associated with several experimental conditions and behaviors.

Local field potentials reflect cortical population dynamics in a region-specific and frequency-dependent manner.

The spiking activity of populations of cortical neurons is well described by the dynamics of a small number of population-wide covariance patterns, whose activation we refer to as 'latent dynamics'. These latent dynamics are largely driven by the same correlated synaptic currents across the circuit that determine the generation of local field potentials (LFPs). Yet, the relationship between latent dynamics and LFPs remains largely unexplored. Here, we characterised this relationship for three different regions of primate sensorimotor cortex during reaching. The correlation between latent dynamics and LFPs was frequency-dependent and varied across regions. However, for any given region, this relationship remained stable throughout the behaviour: in each of primary motor and premotor cortices, the LFP-latent dynamics correlation profile was remarkably similar between movement planning and execution. These robust associations between LFPs and neural population latent dynamics help bridge the wealth of studies reporting neural correlates of behaviour using either type of recordings.

Emerging ideas and tools to study the emergent properties of the cortical neural circuits for voluntary motor control in non-human primates.

For years, neurophysiological studies of the cerebral cortical mechanisms of voluntary motor control were limited to single-electrode recordings of the activity of one or a few neurons at a time. This approach was supported by the widely accepted belief that single neurons were the fundamental computational units of the brain (the "neuron doctrine"). Experiments were guided by motor-control models that proposed that the motor system attempted to plan and control specific parameters of a desired action, such as the direction, speed or causal forces of a reaching movement in specific coordinate frameworks, and that assumed that the controlled parameters would be expressed in the task-related activity of single neurons. The advent of chronically implanted multi-electrode arrays about 20 years ago permitted the simultaneous recording of the activity of many neurons. This greatly enhanced the ability to study neural control mechanisms at the population level. It has also shifted the focus of the analysis of neural activity from quantifying single-neuron correlates with different movement parameters to probing the structure of multi-neuron activity patterns to identify the emergent computational properties of cortical neural circuits. In particular, recent advances in "dimension reduction" algorithms have attempted to identify specific covariance patterns in multi-neuron activity which are presumed to reflect the underlying computational processes by which neural circuits convert the intention to perform a particular movement into the required causal descending motor commands. These analyses have led to many new perspectives and insights on how cortical motor circuits covertly plan and prepare to initiate a movement without causing muscle contractions, transition from preparation to overt execution of the desired movement, generate muscle-centered motor output commands, and learn new motor skills. Progress is also being made to import optical-imaging and optogenetic toolboxes from rodents to non-human primates to overcome some technical limitations of multi-electrode recording technology.