How does the human brain construct a stable and coherent perceptual world amidst rapidly changing visual inputs? Sequential dependence is a key phenomenon for understanding this process: current perception depends not only on immediate input but is also influenced by the sensory content of the previous moment. Interestingly, this influence manifests in two seemingly contradictory forms—sometimes the current perception is "pulled" toward the previous stimulus (attraction effect), and sometimes it is "pushed away" (repulsion effect). Attraction effects often appear in high-level cognitive stages like decision-making, while repulsion effects are common in the early visual cortex. Why do these two effects coexist, and how do their underlying neural mechanisms coordinate?
Research Overview
In a study published at NeurIPS 2025, the research team led by Associate Professor Yuanyuan Mi from the Department of Psychological and Cognitive Science at Tsinghua University and Professor Nihong Chen from the School of Psychology at South China Normal University constructed a two-layer Continuous Attractor Neural Network (CANN) model. This study marks the first time that two synaptic plasticity mechanisms—Short-Term Depression (STD) and Short-Term Facilitation (STF)—have been integrated into a single computational framework to systematically reveal the neural basis of repulsion and attraction in sequential dependence.
Model Mechanisms
The model utilizes a hierarchical structure to simulate different processing stages in the brain:
Low-Level Network (STD-Dominant): Simulates the rapid adaptation characteristics of the early visual cortex. Previous stimuli consume local synaptic resources (STD), causing the neural representation of subsequent stimuli to deviate from the original direction, resulting in repulsion. This repulsion is strongest at short stimulus intervals and decays as STD recovers.
High-Level Network (STF-Dominant): Simulates decision-making brain regions. Memory traces from previous stimuli persist for several seconds, enhancing synaptic efficacy (STF) and biasing subsequent perception toward historical directions, resulting in attraction. This attraction is most significant when intervals are shorter than the STF decay time.
Significance and Bayesian Interpretation
This two-layer model successfully explains sequential dependence effects found in human psychophysical experiments. Within a single trial, repulsion triggered by STD is transmitted from the low-level to the high-level network. Across trials, memory traces maintained by STF in the high-level network continue to influence incoming stimuli, producing attraction.
Furthermore, the study provides a biological implementation path for a two-stage Bayesian inference framework:
Likelihood Weakening: The low-level STD mechanism corresponds to weakening the "likelihood," pushing perception away from previous stimuli.
Prior Enhancement: The high-level STF mechanism corresponds to "prior enhancement," biasing decisions toward historical information.
Conclusion
By constructing a two-stage neural network with short-term inhibition and facilitation, this study unifies repulsion and attraction as emergent properties of the same neural dynamical system at different levels. It clarifies how the brain balances sensitivity to new information with the ability to integrate historical information, opening new directions for understanding the coordination of stability and flexibility in perception-decision circuits.
Authors and Funding
First Author: Xiuning Zhang, Research Assistant in Prof. Yuanyuan Mi’s group.
Corresponding Authors: Associate Professor Yuanyuan Mi (Tsinghua University) and Professor Nihong Chen (South China Normal University).
Key Contributors: PhD student XinCheng Lv (Tsinghua University), Professor Luo Huan, and Assistant Researcher Zhang Huihui (Peking University).
Funding: Supported by the Sci-Tech Innovation 2030 Major Project, the National Natural Science Foundation of China (Key Project and Innovation Research Group Project), and the Guangdong Provincial Center for Brain Science and Human Quality Development.
Paper Link: https://openreview.net/forum?id=rzJkKeliDK
Faculty Profile
Yuanyuan Mi
Title: Associate Professor and Doctoral Supervisor, Department of Psychological and Cognitive Science.
Research Direction: Computational Neuroscience.
