Computing-in-memory architecture for Kolmogorov-Arnold networks based on tunable Gaussian-like memory cells.

Area of Science:

• Neuromorphic Engineering
• Artificial Intelligence

Background:

• Kolmogorov-Arnold networks (KANs) offer advantages over traditional multilayer perceptrons due to their flexible activation functions.
• Hardware implementation of KANs' basis functions is computationally expensive, hindering practical applications.

Purpose of the Study:

• To design a cost-effective hardware architecture for Kolmogorov-Arnold networks.
• To improve the energy efficiency and flexibility of KANs for complex computations.

Main Methods:

• Developed a Gaussian-like memory cell combining a Gaussian transistor and a memristor for tunable current-voltage responses.
• Constructed circuits using these cells to enable parallel inference computation for KANs.

Main Results:

• The proposed architecture successfully implemented KANs for diverse tasks like function regression, image recognition, and time-series forecasting.
• Demonstrated significant improvements in energy efficiency compared to existing methods.
• Validated the algorithmic advantages of KANs in the new hardware.

Conclusions:

• Gaussian-like memory cells provide a promising computing-in-memory solution for Kolmogorov-Arnold networks.
• The developed architecture enhances the flexibility and efficiency of the neuromorphic computing paradigm.