Select Publications

Much effort has been spent clustering neurons into transcriptomic or functional cell types and characterizing the differences between them. Beyond subdividing neurons into categories, we must recognize that no two neurons are identical and that graded physiological or transcriptomic properties exist within cells of a given type. This often overlooked “within-type” heterogeneity is a specific neuronal implementation of what statistical physics refers to as “disorder” and exhibits rich computational properties, the identification of which may shed crucial insights into theories of brain function. In this perspective article, we address this gap by highlighting theoretical frameworks for the study of neural tissue heterogeneity and discussing the benefits and implications of within-type heterogeneity for neural network dynamics, computation, and self-organization.

Long-term sustained pain following acute physical injury is a prominent feature of chronic pain conditions. Populations of neurons that rapidly respond to noxious stimuli or tissue damage have been identified in the spinal cord and several nuclei in the brain. Understanding the central mechanisms that signal ongoing sustained pain, including after tissue healing, remains a challenge. Here we use spatial transcriptomics, neural manipulations, activity recordings and computational modelling to demonstrate that activity in an ensemble of anatomically and molecularly diverse parabrachial neurons that express the neuropeptide Y (NPY) receptor Y1 (Y1R neurons) is increased following injury and predicts functional coping behavior. Hunger, thirst or predator cues suppressed sustained pain, regardless of the injury type, by inhibiting parabrachial Y1R neurons via the release of NPY. Together, our results demonstrate an endogenous analgesic hub at pain-responsive parabrachial Y1R neurons.

Endogenous activity, particularly periodic neural firing, appears across the animal kingdom and can arise through a diverse array of mechanisms. We speculate that periodic neural activity may have been a hallmark of the earliest nervous systems' function, and that modulation of this activity may have thus been an early site of behavioral control.

In contrast to information-encoding units in a computer, neurons are heterogeneous, i.e., they differ substantially in their electrophysiological properties. How does the brain make use of this heterogeneous substrate to carry out its function of processing information and generating adaptive behavior? We analyze a mathematical model of networks of heterogeneous spiking neurons and show that neural heterogeneity provides a previously unconsidered means of controlling computational properties of neural circuits.

A critical barrier to intracortical brain-computer interface (iBCI) adoption is our inability to consistently record the same neurons over long spans of time; this signal drift requires iBCIs to be recalibrated every few days. Here, we show that domain adaptation methods can be used to "align" neural recordings between days, stabilizing iBCI performance for up to several months.

The past decade has seen a proliferation of methods for discovering a vocabulary of animals' actions in a data-driven manner; such methods typically work by recording the spontaneous actions of animals and looking for recurring structure. However, assessing the quality of a learned representation is not easy. In the Multi-Agent Behavior (MABe22) benchmark, we present three large datasets of interacting agents for use in representation learning; each dataset is accompanied by ground-truth labels of animals' states and actions. We propose that a "good" representation of animal behavior is one from which this relevant state and action information can be decoded without having been provided during representation learning.

Next-generation mean-field models are a powerful mathematical tool for modeling the emergent dynamics of spiking neural networks. However, the typical mean-field model treats neurons as identical and interchangeable units, neglecting the biological reality of both distinct neuronal cell types and of variation within a cell type. To address both points, we derive mean-field models for biologically interpretable Izhikevich model neurons and introduce neuronal heterogeneity in the form of distributed spike thresholds, contrasting dynamics of spiking and mean-field models to show where the mean-field assumptions break down.

The ventrolateral portion of ventromedial hypothalamus (VMHvl) plays a central role in regulating within-species aggression. While many VMHvl neurons show complex dynamics that are difficult to interpret, population-level analysis using a switching linear dynamical system reveals a single dimension of VMHvl activity that acts like a leaky integrator: neuronal activity along this dimension grows in magnitude as animals escalate social encounters from investigation to outright attack. Across animals, the time constant of dynamics along this dimension is highly correlated with aggression, suggesting that this time constant might be tuned to adjust an animal's willingness to fight.

Theories and models of behavior vary in their objective, language, and level of mechanistic detail. In this review, we present a delineation of behavior models into broad categories: descriptive models that identify what an agent did, generative models that tell us how an agent can be expected to behave given a set of conditions, and normative models that identify why a behavior occurred by defining the underlying principles that inform its structure.

The development of new algorithms for automated behavior analysis is throttled by a lack of large, high quality datasets that can be used to benchmark algorithm performance: without these datasets, any lab aiming to develop and test a new method must first collect their own training data. Here, we document and release three benchmark datasets for supervised behavior classification, alongside the results of an accompanying online behavior classification competition, the 2021 Multi-Agent Behavior Challenge.

We present the Mouse Action Recognition System (MARS), a computer vision and machine learning platform for markerless pose estimation and supervised behavior classification in socially interacting mice. MARS automatically detects three types of social interaction with human-level accuracy, permitting high-throughput social behavior screening. Accompanying platform BENTO supports analysis and visualization of joint neural and behavioral datasets.