In-Memory Machine Learning Using Hybrid Decision Trees and Memristor Crossbars.

We propose a method to design in-memory hybrid decision tree (HDT) circuits using memristor crossbars. Decision Trees (DTs) are a well known machine learning algorithm that carries multiple benefits when compared to deep neural networks. They are easily interpretable, fast and they require less data to train. These benefits make them a popular choice in wide-ranging applications that include edge devices and particle physics. We propose a HDT coupled with a multilayer perceptron (MLP) for creating a flexible nonlinear decision boundary which leads to better accuracy. Using this approach, we obtain a test accuracy of 0.90 for the MNIST dataset, which outperforms the state- of-the-art (SOTA). We map this decision tree onto crossbars which are purely memristor based. They utilize zero transistor and one memristor per cell and employ sneak-paths for flow- based in-memory computations. Due to the absence of transistors, our designs are radiation degradation resistant, serving their application in radiation-rich environments, and require less switching energy, making them energy efficient.