Implementing in-situ self-organizing maps with memristor crossbar arrays for data mining and optimization

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence
• Materials Science

Background:

• Self-organizing maps (SOMs) are vital unsupervised learning tools for high-dimensional data analysis.
• Hardware implementation of SOMs faces challenges due to complex similarity calculations and neighborhood determination.

Purpose of the Study:

• To experimentally demonstrate a novel memristor-based Self-Organizing Map (SOM).
• To leverage memristor crossbar arrays for efficient hardware implementation of SOM algorithms.

Main Methods:

• Utilized Ta/TaOx/Pt 1T1R memristor crossbar arrays for SOM hardware implementation.
• Employed additional rows in crossbar arrays to directly compute input vector and weight matrix similarities.
• Leveraged the topological structure of the array and physical laws for computation, avoiding complex circuits.

Main Results:

• Achieved significant improvements in computing speed, throughput, and energy efficiency compared to CMOS-based SOMs.
• Successfully demonstrated data clustering, image processing, and solving the traveling salesman problem.
• Validated the feasibility of direct hardware similarity calculation for identifying best matching units.

Conclusions:

• The memristor-based SOM offers a highly efficient and low-power hardware solution for machine learning.
• This work extends the capabilities of memristor-based neuromorphic computing systems.
• Physical implementation of SOMs in memristor crossbar arrays paves the way for advanced AI hardware.