How do we glean concrete insights from neural data?

Our goal is to develop methodologies for obtaining an interpretable model of an organism’s brain as it senses, computes, and acts. Towards this goal I'm focused on developing the interpretation of high-performing functional model (e.g., DNNs), creating estimators for high-dimensional neural response properties to guide data collection and generate theoretical insight, and advancing statistical methods to integrate physiology, perturbations, and neural recordings data for efficient mechanistic model inference. In essence, our work seeks to discover the path from the collection of empirical neural data to a comprehensive understanding of the brain. Below we outline projects and papers that have made concrete steps in these directions.

Mechanistic model inference

A long-standing goal of neuroscience is to obtain a mechanistic model of the nervous system. This would allow neuroscientists to explain animal behavior in terms of the dynamic interactions between neurons. I have proposed a framework for efficiently estimating a dynamical system of the whole fly brain that is not biased by unobserved confounders (e.g., unobserved neurons).

Leveraging anatomical information to efficiently learn a dynamic causal model of the whole fly brain (A) Example experiment with observed neurons (black circles within FOV of camera), a subset are source neurons that express opsin (neurons under red light). (B) To infer mechanistic models I leverage anatomical information from the whole fly brain connectome. (C) It can be used as a prior on model parameters to drastically improve efficiency. (D) Analysis of this prior can propose putative circuits that have strongest effects on dynamics. For example an eigendecomposition of the connectome re-discovers ring neurons in the ellipsoid body (E) simulations of this circuit shows it recapitulates winner take all dynamics.

References

Dean Pospisil, Max Aragon, Jonathan Pillow (2024). From connectome to effectome: learning the causal interaction map of the fly brain. (Nature).

Analysis of complex sensory representations

Decades of work by neuroscientists have resulted in a huge diversity of findings on the nature of sensory representations in the brain. I seek to unify and extend this work with computational models that predict responses to arbitrary sensory inputs and can reproduce results across experiments.

Artiphysiology approach. (A) Prior electrophysiological work found sensory tuning in V4 where neurons encoded the curvature of shape boundaries. For example, responding to a sharp point at the top of a shape. (B) I reproduced this experiment in a deep neural network (DNN) trained for object recognition-- a task V4 is thought to support in the brain. I found the DNN showed shape selectivity indistinguishable from that of the brain. I found this selectivity extended to natural images (for example this artificial unit was most driven by natural objects with a sharp point on top). (C) I have leveraged the accessibility of the DNN to uncover circuits involved in neural like response properties.

References

Estimation of high-dimensional statistical structure

Brain regions contain millions of neurons and their patterns of activity have been termed a ‘population code’. The properties of this code are of critical importance to understanding brain function (e.g., disruption of behavior is typically only seen when large neural populations are damaged). Novel recording techniques have allowed recordings of neural activity from large populations but statistical techniques in neuroscience have lagged behind in addressing the challenges of drawing conclusions from noisy high-dimensional measurements. I am directly addressing this issue to solve concrete problems by building novel estimators

The challenge of estimating eigenvalues from data. (A) Tuning curves of three neurons for a 1-dimensional stimulus (gray traces). (B) Same three tuning curves plotted jointly in a 3D response space. (C) Joint tuning curve centered and plotted along principal axes of variation. (D) Eigenspectrum, which describes the variance along each principal component of the joint tuning curve. (E) Noisy estimates of individual tuning curves at the same three points along the tuning curve (black points). True tuning curve is unknown (light grey trace). (F) Noisy estimate of the joint tuning curve (black dots). (G) Estimated joint tuning curve centered and rotated to align with its principal components; the resulting curve is 2-dimensional, since 3 points defined a plane. (H) Eigenspectrum of the estimated joint tuning curve. Only two eigenvalues are non-zero, and thus later eigenvalue of true tuning curve are missing.

References

Evaluation of model performance

Model evaluation is a neccesary step in building quantitative models of the nervous system. Most current approaches do not account for the pessimistic bias introduced by limited training data and unexplainable noise in measurements.

Estimation of R2 between neural tuning and a mulitvariate linear model (A) Models of sensory tuning often regress a set of computed stimuli features onto (B) average neural responses to stimuli. The correlation between prediction and data underestimates the performance of the model because both the estimate of regression weights and the averaged neural responses are noisy. (C) I have developed estimators for the performance in the limit where the number of repeats goes to infinity (i.e., no neural noise) and (D) when the number of repeats and stimuli goes to infinity (i.e., model parameters are optimal).

References

Pospisil DA, Bair W (2021). The unbiased estimation of fraction variance explained by a model. PLOS Computational Biology.

Estimation of correlation accounting for measurement error

The measurement of correlation between noisy datasets is ubiquitous and plays a critical role in sensory neuroscience. The r2 between two sets of mean neural responses is fundamentally relevant to lines of research that compare tuning curves across neurons to understand functional organization and population encoding and to studies that compare tuning curves within the same neuron across stimulus transformations to understand invariance in cortical representations

Estimating the relationship between neurons accounting for finite samples and measurement noise. (A) Estimates of tuning curves (solid trace) are noisy samples (open circles). Thus the naive estimate of correlation underestimates correlation between tuning curves (here 1). (B-C) I created an estimator that corrects for this bias (blue trace on diagonal) even when the naive estimate is highly biased (orange trace below diagonal). I extended this estimator to the more challenging case when neurons have correlated noise. With these estimators I made novel observations on invariance in visual cortex and signal correlation.

References