Jay Yadawa

Sina Mohseni, Haotao Wang, Zhiding Yu et al.

The open-world deployment of Machine Learning (ML) algorithms in safety-critical applications such as autonomous vehicles needs to address a variety of ML vulnerabilities such as interpretability, verifiability, and performance limitations. Research explores different approaches to improve ML dependability by proposing new models and training techniques to reduce generalization error, achieve domain adaptation, and detect outlier examples and adversarial attacks. However, there is a missing connection between ongoing ML research and well-established safety principles. In this paper, we present a structured and comprehensive review of ML techniques to improve the dependability of ML algorithms in uncontrolled open-world settings. From this review, we propose the Taxonomy of ML Safety that maps state-of-the-art ML techniques to key engineering safety strategies. Our taxonomy of ML safety presents a safety-oriented categorization of ML techniques to provide guidance for improving dependability of the ML design and development. The proposed taxonomy can serve as a safety checklist to aid designers in improving coverage and diversity of safety strategies employed in any given ML system.

Sina Mohseni, Arash Vahdat, Jay Yadawa

Detecting out-of-distribution (OOD) samples plays a key role in open-world and safety-critical applications such as autonomous systems and healthcare. Recently, self-supervised representation learning techniques (via contrastive learning and pretext learning) have shown effective in improving OOD detection. However, one major issue with such approaches is the choice of shifting transformations and pretext tasks which depends on the in-domain distribution. In this paper, we propose a simple framework that leverages a shifting transformation learning setting for learning multiple shifted representations of the training set for improved OOD detection. To address the problem of selecting optimal shifting transformation and pretext tasks, we propose a simple mechanism for automatically selecting the transformations and modulating their effect on representation learning without requiring any OOD training samples. In extensive experiments, we show that our simple framework outperforms state-of-the-art OOD detection models on several image datasets. We also characterize the criteria for a desirable OOD detector for real-world applications and demonstrate the efficacy of our proposed technique against state-of-the-art OOD detection techniques.