Self-supervised learning through the eyes of a child

TL;DR

This paper investigates how self-supervised deep learning models can develop high-level visual representations from natural egocentric videos of children, shedding light on early visual development and the role of innate biases.

Contribution

It demonstrates that generic self-supervised learning on realistic videos can produce high-level visual representations similar to those in children, advancing understanding of visual development.

Findings

Powerful visual representations emerge from natural videos.

Self-supervised learning models can mimic early visual development.

Results support the sufficiency of generic learning mechanisms.

Abstract

Within months of birth, children develop meaningful expectations about the world around them. How much of this early knowledge can be explained through generic learning mechanisms applied to sensory data, and how much of it requires more substantive innate inductive biases? Addressing this fundamental question in its full generality is currently infeasible, but we can hope to make real progress in more narrowly defined domains, such as the development of high-level visual categories, thanks to improvements in data collecting technology and recent progress in deep learning. In this paper, our goal is precisely to achieve such progress by utilizing modern self-supervised deep learning methods and a recent longitudinal, egocentric video dataset recorded from the perspective of three young children (Sullivan et al., 2020). Our results demonstrate the emergence of powerful, high-level visual representations from developmentally realistic natural videos using generic self-supervised learning objectives.