Learning the Synaptic and Intrinsic Membrane Dynamics Underlying Working Memory in Spiking Neural Network Models.

Area of Science:

• Computational Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Recurrent neural network (RNN) models are valuable for studying brain computations.
• Existing models often use continuous signals and neglect spiking neuron properties.
• Understanding intrinsic neuronal parameters is crucial for realistic neural circuit modeling.

Purpose of the Study:

• To develop a spiking RNN model capable of training both synaptic and membrane parameters.
• To investigate how intrinsic neuronal properties and connectivity contribute to cognitive tasks.
• To elucidate the neural mechanisms underlying working memory.

Main Methods:

• Developed a novel method to train synaptic and membrane parameters in spiking RNNs.
• Trained the model on a diverse set of cognitive tasks.
• Analyzed emergent synaptic and membrane parameter dynamics, particularly for working memory tasks.

Main Results:

• The model achieved diverse, task-specific synaptic and membrane parameters.
• Fast membrane time constants and slow synaptic decay emerged for working memory tasks.
• Fast membrane properties were crucial for stimulus encoding, and slow synaptic dynamics for memory maintenance.

Conclusions:

• This approach provides insights into the interplay of connectivity and intrinsic neuronal properties in neural populations.
• Optimized spiking RNNs can reveal how specific neural parameters support cognitive functions.
• The model advances computational tools for neuroscience research.