AI’s Mirror to the Infant Brain: Tracing the Development of Spatial Understanding
A new study by CDS PhD student Guy Davidson sheds new light on the parallels between the developmental stages of infant cognition and the capabilities of modern AI. Davidson’s upcoming paper, “Spatial Relation Categorization in Infants and Deep Neural Networks,” co-authored with CDS Assistant Professor of Psychology and Data Science Brenden Lake and former CDS Research Scientist Emin Orhan, is set for publication in Cognition in early 2024.
This research began with a look back at developmental psychology studies from the late nineties and early 2000s, which revealed that infants begin to categorize spatial relations at three to four months. This means that they start to understand how objects are positioned in relation to each other, such as above, below, or between. This capability is a fundamental aspect of cognitive development, forming the foundation for more complex understanding as they grow.
Davidson and his co-authors’ innovative approach involved comparing these infant capabilities with those of deep neural networks in computer vision. These networks, which had not been specifically trained to categorize spatial relations, were tested to see if they could process images representing these relations and how their performance compared to that of infants. Furthermore, Davidson and colleagues evaluate not only models trained on big data from the web, but also models trained by Orhan on data from an infant head-mounted camera. This dataset, SAYCam, offers a better comparison to the visual input that a child experiences in development.
Like infants, AI models showed a gradation in their ability to process spatial relations. The models more easily categorized simpler relations, such as ‘above vs. below’ (as illustrated below, in 2D (left) and 3D (right)), akin to what infants do at an early developmental stage. As the complexity of the spatial relations increased, both infants and AI models found it more challenging to categorize them accurately, mirroring the developmental progression seen in human infants.
‘Above vs. below’ spatial relation illustrated in 2D (top) and 3D (bottom). AI models’ ability to classify new images as “same” (in this case, “below”) versus “other” (in this case, “above”) resembled that of infants’.
However, there were significant differences between human and machine learning. One notable divergence was that while infants perform equally well in differentiating between orientations like above vs. below and left vs. right, AI models show a notable weakness in differentiating between an object being to the left or right of something. After investigating, Davidson’s team realized that AI models struggle with distinguishing left-right due to training techniques intended to get more mileage out of training data that duplicate images left-right, leading to a loss in directional specificity.
This work is the fruit of a persistent collaboration with Brenden Lake, Davidson’s advisor, that spans the latter’s entire PhD tenure at CDS. This partnership took root in a computational cognitive modeling class during Davidson’s first year, evolving from a final project into a series of scholarly endeavors. Their intellectual journey included an initial submission to a conference workshop, followed by a presentation at a cognitive science conference in 2021, eventually culminating in the manuscript soon to be published. Davidson’s work, under Lake’s mentorship, stands as a testament to the power of academic mentorship and the profound growth that can unfold from the kernel of a single, well-placed academic inquiry.
As Davidson nears the completion of his PhD, his research stands as a vivid example of the synergy between psychology, technology, and AI. This kind of research at CDS underlines the importance of interdisciplinary approaches in unraveling the mysteries of cognition, both in the cradle of human development and in the silicon cores of AI.
By Stephen Thomas
