Endogenously motivated looking in infants
Decades of developmental psychology has shown that infants actively structure their own learning through flexible selection of attentional targets. But the precise rules by which infants look remain elusive. I will first discuss models that describe how infants decide what to look at, and their limitations. One limitation relates to the difficulty of capturing all the regularities that infants are sensitive to when deciding where to look, and I discuss some of our work attempting to close this gap. Then, I argue that there is a second limitation, which results from the assumption that infants' endogenously motivated looking is best described by a single, domain-general objective. I discuss our theoretical work suggesting that infant looking might instead be best described by multiple, domain-specific objectives.
Factorized representations in brains and machines
Learning efficiency and robustness to variations in the environment are hallmarks of intelligence. However, current AI algorithms pale in comparison to humans on measures of learning efficiency and robustness. Compositional representations, namely representations that encode data in terms of recombinable primitives, have been shown to facilitate learning efficiency and robustness in some contexts. I will describe the notion of a factorized representation (a particular type of compositional representation) and will empirically show that factorization correlates with policy robustness in deep neural networks trained on a toy control task. I will then describe a planned long-term project to explore factorization of object features, position, and motion in the primate brain, highlighting the relevance of this project to the binding and visual remapping problems.