Humans possess the remarkable ability to solve complex and novel tasks, whereas artificial neural network models still struggle with some tasks, such as abstract visual reasoning. These tasks require identifying the abstract pattern among different objects, which is natural for humans, for instance, identifying whether two objects are the same or different. While substantial progress has been made in approximating human-like reasoning abilities in such tasks, artificial agents still fail to match the out-of-distribution and sample-efficient generalization shown by humans. In order to improve such abilities, my research focuses on developing artificial neural network models that implement processing mechanisms identified in the study of human cognition and brain function. The talk will be centered around abstract visual reasoning tasks that are often used in standardized measures and laboratory studies of human intelligence. I will first demonstrate the effectiveness of generic object-focused low-dimensional representations for such tasks through a simple model using slot-based unsupervised representation learning along with a Transformer-based reasoning module. Then, I will describe two models integrating object-centric representations with strong inductive biases for relational abstraction, which help us to infer patterns among objects abstracting over the specific details in ways that generalize to completely new objects in abstract visual reasoning. I will show that such models not only achieve human-like out-of-distribution generalization but do so in a highly sample-efficient manner. Finally, I will conclude with some preliminary thoughts on an ongoing project for incorporating object-centricity into large vision language models to further improve their out-of-distribution generalization in visual reasoning.

Adviser: Jonathan Cohen