Automated Controlled-Rearing Methods
Although scientists have pondered the origins of the mind for centuries, progress has been hampered by the lack of high-precision methods for studying newborn organisms in controlled environments. Thus, one of our primary goals is to develop high-precision controlled-rearing methods for studying newborn organisms. We accomplish this through automation. With automation, we can monitor newborn subjects’ behavior continuously (24/7) within strictly controlled virtual environments. By collecting vast amounts of data from each newborn subject, we can study the development of their mental abilities with high precision. Below, we describe the four types of automated controlled-rearing chambers that we have developed in our lab.
Passive Controlled-Rearing Chambers
These chambers allow us to present animations to newborn subjects on a predetermined schedule. This approach enables the study of unsupervised visual learning and spontaneous visual preferences in newborn organisms.
Unsupervised Visual Learning
Much of what we know about the world comes from unsupervised visual learning. Using our passive controlled-rearing chambers, we can study unsupervised visual learning during the earliest stages of learning, when a newborn organism builds their first object concept. The animations above show a sample unsupervised visual learning experiment. During the Input Phase (left), newborn chicks were reared with a single virtual object moving on a single background. During the Test Phase (right), we measured what the chicks had learned from that input, by testing their object recognition abilities across novel viewpoints and novel backgrounds. We found that chicks can build abstract (view-invariant & background-invariant) object concepts from the sparse visual input provided in the Input Phase, showing that the outputs of newborn object recognition generalize far beyond the input coming in through the senses. More generally, this approach allows us to explore how specific visual inputs are transformed into specific behavioral outputs in newborn organisms.
Spontaneous Visual Preferences
Visual preferences influence where an organism looks, and thus, what they learn from the environment. Understanding the origins of perception and cognition therefore requires understanding the nature of early emerging visual preferences. Using our passive controlled-rearing chambers, we can study visual preferences during the earliest stages of learning and chart how those preferences change across the first weeks of life. Two sample visual preference experiments are shown above. From the onset of vision, newborn chicks were reared with two visual stimuli, and we measured whether they had a preference for one stimuli over the other. We found that newborn chicks have early emerging preferences for slowly moving objects (left) and smoothly moving objects (right). This finding is important because newborn chicks build more accurate object concepts from slow and smooth visual object input.
Interactive Controlled-Rearing Chambers
These interactive chambers monitor the subject’s movements and choices and update the environment—in real time—as a function of those behaviors. The system is controlled by a depth-sensing camera connected to a video game engine (Unity). This approach allows us to raise newborn organisms in immersive and interactive virtual environments (“video game worlds”), allowing exploration of a range of abilities. Some sample tasks are shown below.
Reinforcement learning provides a normative account of how agents learn to optimize their control of an environment. Understanding the origins of the mind therefore requires understanding the role of reinforcement learning in the development of perception and cognition. The picture above shows a sample reinforcement learning experiment with newborn chicks. On each trial, the chick is shown two occluders on a display wall. One occluder shows a “correct” choice while the other occluder shows an “incorrect” choice. If the chick moves to the correct occluder, then the occluders drop to reveal the imprinted object (reward). If the chick moves to the incorrect occluder, then the occluders drop to reveal no objects (no reward). After spending 30s with the imprinted object, the imprinted object leaves the chamber and a new trial is presented on the opposite display wall.
Object Search Task
Understanding how object cognition emerges in the brain is a central goal in psychology, neuroscience, and artificial intelligence. These interactive chambers allow us to study the emergence of object cognition and intuitive physics using object search tasks. Two sample object search tasks are shown above. Trials are triggered when the chick is active and alert. The video on the left shows an object permanence trial with object movement, and the video on the right shows an object permanence trial with invisible displacement.