One of the great unsolved mysteries in human cognition and developmental science concerns the origins of the mind. How do newborns transform streams of sensory input into knowledge? What mechanisms underlie newborn perception and cognition? Despite widespread interest in these questions, two methodological barriers have hindered progress. First, it is not possible to conduct controlled-rearing studies on newborn humans, so researchers could not perform causal tests of how specific experiences shape the newborn mind. Second, newborn humans cannot be tested for long periods of time, limiting our ability to study newborn cognitive development with high precision.
To overcome these barriers, my lab developed an automated controlled-rearing method with a newborn animal model: the domestic chick. This method allows us to study newborn subjects continuously (24/7) within strictly controlled virtual environments. By recording each subject’s behavior continuously, we can achieve low measurement error and obtain large, robust effects across experiments. With massive amounts of data from each subject, we can also examine whether each newborn succeeded or failed at an experimental task and study individual differences across newborn subjects.
My lab is currently using automated controlled rearing to study how object recognition emerges in newborn brains. With new funding support, we will extend this approach to other domains, including face recognition, action perception, naïve physics, and numerical cognition.
To understand the development of these abilities, we will focus on three interconnected components of newborn minds:
In parallel to the controlled-rearing experiments, my lab will build artificial newborn agents. These agents will learn about the world autonomously, using a combination of information seeking, sensory processing, and reinforcement learning systems. To measure the learning abilities of the agents, we will construct virtual worlds that simulate our controlled-rearing experiments. We will then test whether the artificial agents learn about the worlds in the same way as the newborn subjects tested in our controlled-rearing experiments.
Ultimately, a central goal of my lab is to build a pixels-to-actions artificial agent that learns about the world like a newborn. Building an artificial agent that learns like a newborn would represent an effective measure of success in terms of understanding cognitive development and be a major achievement for the field. More generally, this two-pronged approach—linking controlled rearing to artificial intelligence—provides a unified framework for studying newborn minds.
To promote the transparency, reproducibility, and replicability of controlled-rearing research, my lab will construct an open-source repository containing all of the stimuli and data from our experiments. This repository will be a valuable resource for comparing the learning abilities of newborn animals and artificial agents. Our goal is to make our input-output maps widely available to the scientific community and foster interactions between developmental science, computational neuroscience, and artificial intelligence.