Designing in-Memory Random Forest Memristor Crossbar Circuits

Ensemble decision tree models, such as random forests, have proven to be highly effective and often outperform other machine learning algorithms, including neural networks, on non-specialized tasks. This paper presents a novel approach to creating energy-efficient random forest circuits by utilizing in-memory computing with memristor crossbars. The inmemory PATH-based memristor crossbar provides state-of-theart (SOTA) energy efficiency for decision trees and is expected to deliver SOTA performance for random forests as well. However, no existing method combines and constructs ensembles for random forest implementations. We propose a new method for ensemble construction by converting the one-hot encoded predictions of decision trees into binary values, followed by a majority voting circuit to aggregate the results. The resulting random forest circuit achieves more than 3x improvement in energy efficiency compared to the current SOTA.