TREAM: A Tool for Evaluating Error Resilience of Tree-Based Models Using Approximate Memory

Mikail Yayla, Zahra Valipour Dehnoo, Mojtaba Masoudinejad, Jian-Jia Chen

Abstract: Approximate memories reduce the resource demand of machine learning (ML) systems at the cost of bit errors. ML models have an intrinsic error resilience and are therefore suitable candidates to use with approximate memories. Although the error resilience of neural networks has been considered in many studies, tree-based applications have received less attention. In addition, there is no tool available to specifically evaluate the error resilience of tree-based models. In this work, we present TREAM, a general tool built upon the sklearn framework for injecting bit flips during the inference of tree-based models. TREAM is capable of injecting bit flips into the tree and input parameters, i.e. feature and split values, in addition to feature and children indices. It can also be used for both floating point and integer values.