Encoding Time in Feedforward Trajectories of a Recurrent Neural Network Model.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Systems Neuroscience

Background:

• Brain activity unfolds over time, forming neural dynamics crucial for sensorimotor processing, behavior, and memory.
• Understanding the temporal aspect of neural dynamics is key to deciphering brain function.
• Previous studies suggest sequential neuronal activation encodes time, but the network architecture (feedforward vs. recurrent) and structural role in timing remain unclear.

Purpose of the Study:

• To investigate how recurrent neural networks (RNNs) generate timed neural activity and motor patterns.
• To explore the role of network structure and connectivity in neural timing.
• To model and explain Weber's law within a computational framework.

Main Methods:

• Developed a recurrent neural network (RNN) model with distinct excitatory and inhibitory units.
• Analyzed network connectivity, efficiency, and the balance of excitation (E) and inhibition (I).
• Simulated the generation of multiple timed motor patterns and assessed adherence to Weber's law.

Main Results:

• A single RNN autonomously produced multiple functionally feedforward trajectories, encoding timed motor patterns up to several seconds.
• The model successfully accounted for Weber's law, a key characteristic of timing behavior.
• Network efficiency decreased with an increasing number of stored trajectories, and the E/I balance shifted toward excitation during activation periods.

Conclusions:

• Recurrent network architectures can generate multiple functionally feedforward neural activity patterns through dynamic shifts in the E/I balance.
• The findings suggest that sequential neural activity and Weber's law in timing behavior are explained by recurrent network dynamics.
• The study provides a computational model for understanding the neural basis of timing and motor control.