Fig. 1: Comparison of memristor-based MLPs and GMC-based KANs. | Nature Communications

Fig. 1: Comparison of memristor-based MLPs and GMC-based KANs.

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

Fig. 1: Comparison of memristor-based MLPs and GMC-based KANs.The alternative text for this image may have been generated using AI.

a The structure diagram of MLPs shows the calculation principle of the dense layers and the electrical characteristics of the memristors. b The structure diagram of KANs shows the calculation principle of learnable activation functions and transfer characteristics of GMCs. c Each GMC consists of a Gaussian transistor and a Gr/CIPS/Gr memristor. d Schematic diagram of the device structure and drain current (\({I}_{{\mbox{D}}}\)) -gate voltage (\({V}_{{\mbox{g}}}\)) curve of a Gaussian transistor. Sharp increases and decreases in drain current can be observed by switching between P-type and N-type operations through controlling the operating state of the gate. e The schematic diagram of tunable GMC is realized by stimulating and changing the conductance of the memristor to control the voltage at both ends of the Gaussian transistor, thus realizing the controllable Gaussian-like functions.

Back to article page