The recurrent temporal restricted Boltzmann machine captures neural assembly dynamics in whole-brain activity

Area of Science:

• Computational Neuroscience
• Machine Learning
• Animal Behavior

Background:

• Animal behavior exhibits a mix of random exploration and directed actions driven by neural activity.
• Previous work used compositional Restricted Boltzmann Machines (cRBM) to identify neural assemblies in zebrafish brain activity, capturing stochasticity.

Purpose of the Study:

• To develop a combined stochastic-dynamical model for neural activity.
• To improve the representation of animal behavior by incorporating temporal dynamics.
• To evaluate the efficacy of the recurrent temporal RBM (RTRBM) for analyzing large-scale neural data.

Main Methods:

• Extended compositional RBM (cRBM) to recurrent temporal RBM (RTRBM) using transfer learning.
• Analyzed both simulated and experimental larval zebrafish neural data.
• Estimated multiple RTRBMs at varying temporal resolutions to identify assembly dynamics timescales.

Main Results:

• RTRBMs effectively capture temporal weights in hidden units, representing neural assemblies.
• The temporal RTRBM model demonstrated superior generalization error over the stochastic-only cRBM.
• Accurate representation of neural activity moments in time was achieved.
• Identified distinct timescales of neural assembly dynamics.

Conclusions:

• RTRBMs offer a powerful approach for modeling combined stochastic and time-predictive dynamics in large neural datasets.
• This method advances the understanding of neural mechanisms underlying animal behavior.
• RTRBMs provide a valuable tool for analyzing complex biological systems.