Robust Working Memory in an Asynchronously Spiking Neural Network Realized with Neuromorphic VLSI

Area of Science:

• Neuroscience
• Computer Engineering
• Artificial Intelligence

Background:

• Spiking neural networks (SNNs) are biologically inspired computational models.
• Neuromorphic very-large-scale integration (VLSI) hardware aims to emulate brain function.
• Attractor dynamics are crucial for cognitive functions like memory and decision-making.

Purpose of the Study:

• To demonstrate bistable attractor dynamics in a neuromorphic VLSI hardware implementation of an SNN.
• To investigate the properties of sustained high-firing states and their stability.
• To explore the network's ability to retain information and correct corrupted states.

Main Methods:

• Implementation of a three-population SNN (two excitatory, one inhibitory) using leaky integrate-and-fire (LIF) neurons on neuromorphic VLSI hardware.
• Engineering strong synaptic self-excitation in one excitatory population to create meta-stable states.
• Analyzing evoked and spontaneous transitions between high- and low-firing states, and the network's response to external stimuli and internal fluctuations.

Main Results:

• Successful demonstration of bistable attractor dynamics, sustaining meta-stable high- and low-firing states.
• Evidence of working memory, where the network retains stimulus information after release.
• Observation of spontaneous high-firing states persisting for seconds, significantly longer than circuit timescales.
• Characterization of a continuum between evoked and spontaneous transitions with variable latencies.
• Demonstration of the network's ability to correct corrupted high states within a basin of attraction.

Conclusions:

• Neuromorphic hardware can effectively realize complex attractor dynamics with small neuron populations.
• The implemented SNN exhibits robust working memory and long-term state stability.
• This work highlights the potential of neuromorphic computing for advanced cognitive functions.