Memristor learning model

Built memristor model and simulated learning circuit with python.
Inspired by Spike learning method using STDP(Spike-Time dependency plasticity) material, which is called memristor.

Goal

In this repository, we aimed to build a model to classify MNIST dataset by using Memristor grid model.
Composing memristors into a crossbar architecture, It is actually similar to the layer used in Artificial Neural Network.

SADP (with a reverse circuit)

Since memristance changes according to the amplitude of input-spike,
model could learn things by using SADP(Spike-Amplitude dependency plasticity).
Besides, by simply adding a reverse-input memristor crossbar, we achieved certain stability of the output.

Output

Tried out single epoch with 20000 MNIST data,
and got about 50~60% accuracy.

Memristance(conductivity) distribution after learning process

STRIP 2018 Seminar

For more details, you could refer to this pdf I used for presentation below.

• PDF used in presentation
• STRIP blog