Theme 1: Neuro-inspired Algorithms and Theory

Abstract:

Out-of-Distribution (OoD) inputs are examples that do not belong to the true underlying distribution of the dataset. Research has shown that deep neural nets make confident mispredictions on OoD inputs. Therefore, it is critical to identify OoD inputs for safe and reliable deployment of deep neural nets. One of the key challenges with OoD detection is handling new unlabelled data and distribution shifts especially in an online learning scenario. In this paper we propose Norm-Scaling to address these problems. The proposed Norm-Scaling method applies Z-score normalization over time to normalize the logits separately for each class. This ensures that a single value consistently represents similar uncertainty for various classes. Since the normalization is updated over time it accounts for distribution shifts.  Further, most existing methods apply a single threshold on a similarity  metric to detect OoD inputs from In-Distribution (ID) data. However, we observe that different classes in the dataset have different thresholds to detect OoD data. Thus, applying a single threshold for all classes is not ideal. We show that the proposed method also addresses this issue and outperforms a multi-threshold detector.  Our experiments show that norm-scaling achieves 9.78% improvement in AUROC, 5.99% improvement in AUPR and 33.19% reduction in FPR95 metrics over previous state-of-the-art methods.

 

Bio:

Deepak is currently a Research Assistant at Nano(neuro)electronics Research Laboratory advised by Prof. Kaushik Roy. He received his B.E. in Electronics Engineering from Ramaiah Institute of Technology, India, in 2016, and M.S. degree in Electrical and Computer Engineering from Purdue University, in 2019. He was awarded the bronze medal for academic excellence at Ramaiah Institute of Technology. He is also one of the recipients of College of Engineering scholarship and the ECE summer research grant. After he earned his undergraduate degree, he worked at National Instruments R&D, where he developed signal acquisition and processing frameworks. Later, he developed machine learning and deep learning pipelines at Microsoft. He has also previously been a graduate lecturer at Purdue where he has taught both undergraduate and graduate courses. His current research interests include Deep Learning algorithms with a focus on robustness.