Research

The main goal of Zhang's lab is to address how neural circuit dynamics give rise to neural codes to represent and process uncertain information, and how these produce perception, cognition, and behavior. Below we introduce our research from a computational perspective, each of which can be linked to concrete neurobiology experiments.

Neural circuit implementation of Bayesian (causal) inference

The brain performs Bayesian (causal) inference to interpret the world and to infer the causal relations in the world. Bayesian (causal) inference underlies a wide spectrum of perceptual and cognitive processes, learning, and memory. It is a fundamental question in computational neuroscience how neural circuits implement Bayesian (causal) inference, and how neurons represent uncertainty information. We have conducted a series of theoretical studies along this line. Using multisensory cue integration as an example of Bayesian inference, we proposed that cue integration can be achieved in a decentralized circuit composed of several coupled brain areas (Zhang et al., J. Neurosci., 2016; NeurIPS 2013). We also identified that opposite neurons found in multisensory areas are able to perform casual inference in information integration (Zhang et al., eLife 2019; NeurIPS 2019, 2016). Recently we have been studying how the Poisson spiking variability of cortical neurons drives distributed sampling-based Bayesian inference in cortical circuits (Zhang, Nat. Comms 2023), and how canonical recurrent neural circuits can implement and switch among several Bayesian sampling algorithms (Dong, NeurIPS 22; Sale, NeurIPS 24).

Neural circuit dynamics

Neuronal responses are dynamic and have large response variability. It is important to study how interactions between neurons give rise to these dynamical and variable responses. We are particularly interested in two canonical circuit dynamics used in computational neuroscience: One is Continuous Attractor Neural Network which has been widely used to explain the processing of continuous stimulus (Zhang et al., 2016; 2012); Another is the Excitation-Inhibition balanced spiking network which internally generates large response variability. We are combining the main characteristics of the two canonical networks to study how the network utilizes the structure of recurrent connectivity and its internally generated variability to implement stochastic computation such as sampling-based Bayesian inference.

Invariant and equivariant neural representation

Symmetry (invariance and equivariance) is ubiquitous in the universe and is the foundation of physical laws. Indeed, theoretical physicists have used symmetry to generate and unify physical laws, e.g., unifying four fundamental forces, general relativity, etc. Since the brain stores an internal model of the symmetric world, we hypothesize that symmetry can be used as a designing principle of information processing in neural circuits in the brain. Recently we started an ambitious attempt where we start from the Lie group symmetry to derive concrete, biologically plausible neural circuits in the brain. We have used the 1D translation group to frame the compass neural circuit in fly's brain (Zhang et al., NeurIPS 22), study how recurrent circuits implement temporal scaling equivariance (Zuo et al., NeurIPS 23), and formulate the motion planning as a Lie group operator search problem (Zuo, NeurIPS 24). We believe the group symmetry structure can be used to interpret neural circuit architecture in the brain.

Other research directions

Biological learning and synaptic plasticity.
Neural data analysis.