A programmable triangular neighborhood function for a Kohonen self-organizing map implemented on chip

Area of Science:

• Integrated Circuit Design
• Artificial Neural Networks
• Hardware Acceleration

Background:

• Self-Organizing Maps (SOMs) are crucial for unsupervised learning and data visualization.
• Existing SOM implementations often face limitations in speed and power consumption, especially in hardware.
• Neighborhood functions, like the Gaussian Neighborhood Function (GNF), are key components influencing SOM performance.

Purpose of the Study:

• To introduce an efficient transistor-level implementation of a flexible, programmable Triangular Function (TF).
• To utilize this TF as a Triangular Neighborhood Function (TNF) for ultra-low power SOMs realized as Application-Specific Integrated Circuits (ASICs).
• To demonstrate that the proposed TNF approximates the GNF effectively while offering simpler hardware implementation and improved speed.

Main Methods:

• Transistor-level implementation of a programmable Triangular Function (TF).
• Integration of the TF as a Triangular Neighborhood Function (TNF) within a larger neighborhood mechanism for SOMs.
• Extensive simulations using software models and Hspice for transistor-level verification under various conditions (e.g., supply voltage, temperature).

Main Results:

• The TNF provides a hardware-efficient approximation of the GNF.
• The parallel implementation achieves rapid distance calculations (≤11 ns) and output value computation (≤6 ns), enabling full map adaptation in ≤17 ns.
• Ultra-low power SOMs realized with this TNF are significantly faster (hundreds of times) than PC-based implementations.
• Low signal resolutions (3-6 bits) for the TNF output do not compromise the SOM's learning abilities.

Conclusions:

• The proposed transistor-level TNF offers a highly efficient and fast solution for ultra-low power SOMs in ASICs.
• This approach simplifies hardware implementation compared to GNF while maintaining performance.
• The design demonstrates robustness across different operating conditions and signal resolutions, making it suitable for commercial applications.