Designing Energy-Efficient PATH-based Decision Tree Memristor Crossbar Circuits

We propose an approach to designing energy-efficient in-memory decision tree (DT) machine-learning circuits using memristors and PATH-based computing. DT machine learning (ML) algorithms are computationally simple, and they require fewer data to train compared to other ML algorithms, making them an attractive option for edge computing. Existing work on creating energy-efficient DT circuits has utilized FLOW-based computing. However, these approaches require costly write operations before every inference, leading to high energy utilization. In this paper, we propose an algorithm to generate PATH-based DT in-memory crossbar circuit designs that only require a write during circuit synthesis and not before every inference, leading to highly energy-efficient inference. We test the performance of our PATH-based design on multiple standard machine learning datasets and demonstrate that our approach achieves a 47% reduction in energy consumption, on average, compared to the state-of-the-art.