ePoster

An emergent low-rank neural architecture for manual interception of moving targets

Yating Liu, Siqi Li, Yongxiang Xiao, He Cui, Ni Ji
COSYNE 2025(2025)
Montreal, Canada

Conference

COSYNE 2025

Montreal, Canada

Resources

Authors & Affiliations

Yating Liu, Siqi Li, Yongxiang Xiao, He Cui, Ni Ji

Abstract

From catching a baseball intercepting an opponent at karate, movement control in a dynamic setting often requires motor planning in prediction of future target motion. To understand how the brain to form anticipatory motor plan, we trained a modular recurrent neural network model to output similar hand trajectories as macaque monkeys in a manual interception task. In the network modules designed to model the dorsal premotor (PMd) and primary motor (M1) cortices, we observed that rotational neural dynamics during motor preparation transitioned into a subspace encoding hand positions during execution. Similar to neural data from monkey PMd and M1, the hand endpoints can be decoded well before movement onset in both the PMd and M1 modules, suggesting that an anticipatory motor plan emerged as early as in PMd. To reveal the neural network architecture underlying anticipatory planning, we performed cluster analysis on the connectivity within and between network modules. We found four clusters emerged from the M1 module, with strong positive connections within clusters and negative to weak connections between clusters. Surprisingly, the output weights of the M1 clusters mapped to opposite directions along the Euclidean axes, and clusters that mapped to opposing directions mutually inhibited each other. Clustering the PMd units by their projection weights to M1 revealed five clusters, from which opposite projection patterns to antagonizing M1 clusters are consistently found. Cluster-level activity analyses revealed that M1 clusters preferentially synchronized during execution and exhibited motion-invariant tuning to reaching direction. In contrast, PMd clusters exhibited phase-lagged sinusoidal dynamics that transitioned smoothly from preparation to execution. Similar low-rank network architecture and cluster dynamics emerged under alternative model architecture with PMd and M1 modules only. Our findings revealed a potentially general network architecture that converts instantaneous target motion into forward prediction of upcoming movement.

Unique ID: cosyne-25/emergent-low-rank-neural-architecture-e790bf70