Sparse Coding Using the Locally Competitive Algorithm on the TrueNorth Neurosynaptic System

Area of Science:

• Computational Neuroscience
• Neuromorphic Engineering
• Artificial Intelligence

Background:

• Sparse coding is crucial for efficient neural signal representation.
• The Locally Competitive Algorithm (LCA) offers a biologically plausible model for sparse coding.
• Neuromorphic systems like IBM TrueNorth aim to mimic brain functionality.

Purpose of the Study:

• To implement the LCA algorithm on the IBM TrueNorth neurosynaptic system.
• To investigate the feasibility and performance of LCA on neuromorphic hardware.
• To analyze data structures, functional units, and micro-architectural designs for LCA on TrueNorth.

Main Methods:

• Mapping the LCA algorithm onto the TrueNorth architecture.
• Designing data structures and functional processing units for vector-matrix multiplication and non-linear thresholding.
• Implementing dynamical iterative algorithms within the micro-architecture.
• Conducting experimental comparisons using fixed-point arithmetic on TrueNorth versus floating-point on general-purpose computers.

Main Results:

• Successful implementation of LCA on the IBM TrueNorth system.
• Favorable comparison of LCA performance using limited-precision fixed-point arithmetic against standard floating-point computations.
• Analysis of the algorithm's scalability within TrueNorth's constraints.

Conclusions:

• The LCA algorithm can be effectively implemented on neuromorphic hardware like IBM TrueNorth.
• Neuromorphic implementation offers a viable and efficient approach for sparse coding.
• This work demonstrates the potential of brain-inspired computing for complex algorithms.