Developmental Machine Learning

Children are phenomenal learning machines: their ability to generalize well from very few exposures has been an inspiration for the developmental science community. Although the human brain does not behave like machine learning algorithms, utilizing machine learning tools to attempt developmental learning problems in realistic scenarios can help us gain novel insights into mechanisms of child development. On the other hand, it is meaningful to the computer vision community to tackle learning problems in scenarios that resemble human learning in order to devise artificial intelligent agents that can successfully operate in the real-world. While developmental learning intersects with machine learning in many ways, it is important to note that current machine learning methods do not possess the key characteristics of infant learning. We propose to develop an approach to computationally testing hypotheses and findings in developmental psychology using simulation environments that generate the types of perceptual inputs exploited by children during learning, combined with deep learning-based computer vision approaches to quantify learning performance computationally. We investigate the following research directions: 1) Continual Learning from Self-Generated Object Views 2) Generalization in 3D Shape Learning and its Relation to Visual Object Categorization.