cv

Post Doctoral Researcher

Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University

PhD

Fudan University, Shanghai, China

Applied Mathematics

MSc

University of Warwick, Coventry, UK

Computer Science

BSc

Northwestern Polytechnical University, Xi'an, China

Environmental Engineering

Stochastic Forward-Forward Learning through Representational Dimensionality Compression

NeurIPS

The Forward-Forward (FF) algorithm is a biologically plausible alternative to backpropagation (BP) for training neural networks. However, existing goodness functions used in FF neglect the correlated variability between neurons. We proposed a novel goodness function based on dimensionality compression that incorporates second-order statistical structure. Our approach promotes structured representations without the need for negative samples and achieves competitive performance compared to other non-BP methods.

Learning and Inference with Correlated Neural Variability

PNAS Nexus

We developed a new kind of neural network known as the moment neural network (MNN), which depicts how the mean firing rate and the noise correlation of spiking neurons propogate through the network. We show that the MNN can be used to approximate the dynamics of a spiking neural network (SNN) with a high degree of accuracy. As a result, the MNN is a promising theorectical tool to understand how SNN perform probabilistic computing.

Toward a Free-Response Paradigm of Decision-Making in Spiking Neural Networks

Neural Computation

Speed-Accraucy Trade-Off (SAT) is a fundamental property of decision-making in the brain, where the reaction time reflect an agent's uncertainty about the decision. We explained the SAT in the context of spiking neural networks with the help of MNN, and showed that an SNN learning to shape its decision confidence leads to shorter reaction time. By setting a reasonable stopping policy, the SNN can achieve the same performance with much shorter latency in average.

Learning to integrate parts for whole through correlated neural variability

PLOS Computational Biology

For historical reasons, the mean firing rate is throught to be the parimary information carrier in the brain. However, recent studies have shown that the correlated neural variability, which is the trial-to-trial fluctuations in the neural responses, can also carry information. We demonstrated that the correlated neural variability can be used to integrate parts for whole, which is a fundamental computation in the brain.

Moment Neural Network (MNN)

MNN is a new class of deep learning architecture which naturally generalizes rate-based neural networks to second order statistical moments. Once trained, the parameters of the MNN can be directly used to recover the corresponding SNN without further fine-tuning. The trained model captures realistic firing statistics of biological neurons including broadly distributed firing rates and Fano factors as well as weak pairwise correlation.

• Brain-inspired Intelligence
• Neural Coding
• Probabilistic Computing
• Noise Correlation