Research

For the most up-to-date publication list, please visit my Google Scholar page.

The emergence of abstract representations

We show that training feedforward neural networks to perform multiple tasks (here, linear and somewhat nonlinear classifications of a set of latent variables) leads to the emergence of representations that provide good generalization properties for novel tasks (i.e., abstract representations).

Robustness in neural codes

We show that neural representations can be made more reliable by using the higher dimensional codes produced by nonlinear mixed selectivity. We further show that these benefits hold for reasonable numbers of neurons and that nonlinear mixed selectivity is present in the lateral intraparietal area when it is not necessary to support flexible behavior, but would increase robustness.

Coding difficulties for multiple stimuli

At a given time, the brain is receiving multiple different kinds of information about multiple different objects that are out in the world. How does it make sense of it all?

Within a single brain region, animals must bind the different sensory features that arise from a single object together -- this is referred to as the binding problem. In collaborative work with Justin Fine and Ben Hayden, we argue that the reward value associated with offers in a risky gambling task performed by monkeys is bound to its spatial location through the use of semi-orthogonal subspaces. In particular, we show that the different neural subspaces used to encode the value of left and right offers are sufficiently orthogonal to produce reliable binding and, later, decoding of offer value. However, we also show that the subspaces are parallel enough to allow for some generalization of the code for value from one spatial location to another. We argue that this commonality may be important for efficient learning.

Across different brain regions, the animal must integrate distinct representations of the same set of objects. This is referred to as the assignment problem. For example, in the case of integrating sights with sounds, both sensory systems develop an estimate of the spatial position of the source: we show that this redundancy can be used to correctly assign the two sets of percepts to each other. We also show how likely it is to fail, and how changing the amount and kind of overlapping information changes the reliability of assignment.

Directed and undirected tasks

We compared recordings from the primate lateral intraparietal area (LIP) during directed and undirected visual stimulus selection. While the tasks do sharply differ in reward contingencies and training history (the directed task has contingent rewards and was learned by extensive training, the undirected task has no reward contingency and was not trained), they both center on the same action: a saccade selecting a natural image stimulus for overt attention. We show that LIP appears to be preferentially involved in the directed task. This suggests that LIP's role in stimulus selection may be shaped by training and reward contingency.

Reproducibility

We attempted to replicate the findings from a sample of 100 previously published studies in the psychology literature -- and from this generated an estimate of the replication rate of psychological science. I performed one of these replications (with Rebecca Saxe) as a member of the Open Science Collaboration.