Go to ICLR 2024 Conference homepageInvestigating Feature Alignment Under An Infant-Inspired Visual Distribution Shift
Recent work on visual learning in people finds that human infants often experience extended bouts of experience with a small number of familiar objects (e.g., toy ducks at home), with a very long tail of less frequent exposures to less familiar objects (e.g., real ducks at the park). When facing this type of distribution shift between toy ducks and real ducks, learners trying to build coherent representations that bridge these two distributions can leverage at least two distinct types of learning signals: (1) categorical learning signals, which explicitly assign two inputs to the same class (e.g., hearing adults label both toy ducks and real ducks with the same word, ``duck;'' and (2) perceptual learning signals, which implicitly assign two inputs to the same class because of perceived similarities (e.g., both toy ducks and real ducks have bills, wings, and webbed feet). In this paper, we examine how these two types of learning signals interact to impact a learner's cross-domain classification performance, through the lens of feature alignment as an interim goal for the learner. We propose new cluster-based metrics to quantify feature alignment in an infant-inspired two-domain learning problem, and we describe a series of experiments that systematically vary these learning signals to observe impacts on feature alignment and overall learning outcomes.