Modified self-organizing feature map algorithms for efficient digital hardware implementation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Kohonen's self-organizing feature map (SOFM) algorithm is a key unsupervised learning technique.
• The original SOFM algorithm updates weights after each input vector presentation, limiting parallelization.
• Efficient parallel implementation of SOFM is crucial for large-scale datasets and high-performance computing.

Purpose of the Study:

• To develop and analyze variants of the SOFM algorithm optimized for parallel processing.
• To investigate the theoretical properties and convergence of these parallel SOFM variants.
• To demonstrate the practical applicability and performance of the proposed variants on real-world data.

Main Methods:

• Developed two batch-update variants of the SOFM algorithm.
• Provided theoretical proofs for the absorbing class property, ordering, and convergence of 1-D maps.
• Conducted simulations using 2-D maps on 12-D speech data.
• Described a massively parallel hardware system for efficient SOFM implementation.

Main Results:

• Proved that weight configurations form absorbing classes in both variants for 1-D maps.
• Demonstrated convergence and ordering properties consistent with the original SOFM algorithm.
• Showcased that one variant achieves performance comparable to the original SOFM on speech data.
• Validated the practical utility of the parallel variants through a hardware implementation.

Conclusions:

• The proposed parallel SOFM variants maintain the desired properties of the original algorithm.
• These variants enable finer-grained parallelism, suitable for massively parallel systems.
• The optimized variant offers a practical and efficient solution for large-scale SOFM applications.