Emergent Auditory Feature Tuning in a Real-Time Neuromorphic VLSI System

Area of Science:

• Neuromorphic Engineering
• Computational Neuroscience
• Artificial Sensory Systems

Background:

• Ecologically important sounds often have time-varying spectral characteristics.
• Existing neuromorphic auditory models primarily focus on static patterns, treating dynamic patterns as sequences.
• Recent models suggest dynamic feature sensitivity emerges from exposure to stimuli in spiking neural networks.

Purpose of the Study:

• To demonstrate the efficient implementation of a dynamic feature-sensitive neural network in neuromorphic hardware.
• To address challenges in multi-chip asynchronous communication and hybrid analog/digital VLSI realization for neural computation.
• To develop adaptive, neurobiologically plausible, spike-based artificial sensory systems.

Main Methods:

• Modeled a spiking neural network based on thalamo-cortical architecture, incorporating lateral/recurrent connections, axonal delays, and Spike Timing Dependent Plasticity (STDP).
• Implemented the model in hybrid analog/digital VLSI technology for neuromorphic hardware.
• Focused on real-time event-based asynchronous communication for multi-chip systems.

Main Results:

• Developed a functional hardware neural network capable of real-time learning.
• The system demonstrated preferential responses to specific spectro-temporal patterns after exposure.
• Successfully addressed challenges in hardware implementation for dynamic stimulus processing.

Conclusions:

• The implemented neuromorphic hardware effectively processes dynamic spectro-temporal patterns.
• This work represents a significant step towards adaptive, biologically plausible, spike-based artificial sensory systems.
• The availability of this hardware facilitates further development in real-time adaptive sensory processing.