Currently, we are working to build a mechanistic understanding of how the primate brain develops and computes high level neuronal representations of objects and other latent content in the world from visual input, and how those representations support behaviors such as object categorization, identification, and value judgement. Concretely, we empricially study how biolgical neural networks in the visual system transform the pattern of light striking the eyes from initial, pixel-like neural population activity patterns, to new, remarkably powerful forms of neural population activity patterns – neural “representations” that can support our seemingly effortless ability to solve visual intelligence tasks in the real world. And we use those empirical neural and behavioral results and data to guide the construction of scientific, computable models of how those biolgical neural network carry out the critical computations. To approach this complex set of interlocking scientific problems, the work of our research group is directed along several main lines:
Elucidating Neuronal Object Codes
One key direction is to experimentally measure and analyze the patterns of neuronal spiking activity (“codes”) found at the highest levels of the ventral visual stream (primate inferior temporal cortex, IT). At this high level, those neuronal codes have solved the “invariance” problem [3],[4]. While one should not be surprised that such codes exist in the brain, their discovery and continued deeper understanding enables us to focus on the algorithms that construct the codes.
The Quest for Underlying Algorithms
Discovering the key algorithms requires a tight interplay between experiment and theory. For example, we recently discovered that the key invariance properties of neuronal object codes are plastic and can be built from unsupervised, natural visual experience. To explore the potential power of such ideas, we and our collaborators implement and screen large families of brain-constrained models and test them on real-world problems. More generally, we are building a systematic foundation to bring together neuronal data, mechanistic models, and human recognition performance.
The Circuits that Implement those Algorithms
Clever computational algorithms do not exist in a vacuum, but must be implemented in specific neuronal circuits in the brain tissue. We employ high resolution MR and fMRI imaging, microfocal stereo x-ray methods, and optogenetic tools to understand the spatial layout of those circuits in the ventral visual cortex. This information will provide clues about the algorithms at work. It will also allow us to interact with those neuronal circuits to both test hypotheses and potentially enable new brain machine interfaces.