HMI Data AI and Society Seminar

HUMANISING MACHINE INTELLIGENCE - DATA, AI & SOCIETY PUBLIC SEMINAR

Seminar Title: The Familiarity Hypothesis: Explaining the Behavior of Deep Open Set Methods

Abstract: In many applications, object recognition systems encounter objects belonging to categories unseen during training. Hence, the set of possible categories is an open set. Detecting such "novel category" objects is usually formulated as an anomaly detection problem. Anomaly detection algorithms for feature-vector data identify anomalies as outliers, but outlier detection has not worked well in deep learning. Instead, methods based on the computed logits of object recognition networks give state-of-the-art performance. This talk proposes the Familiarity Hypothesis that these methods succeed because they are detecting the absence of familiar learned features. This talk will review evidence from the literature and from our own experiments that supports this hypothesis. It then experimentally tests a set of predicted consequences of this hypothesis that provide additional support. The talk will conclude with a discussion of whether familiarity detection is an inevitable consequence of representation learning and concludes that we can go beyond familiarity detection if we can learn to represent objects in terms of disentangled attributes that support outlier detection.