Research
Understanding the Brain's Fundamental Cognitive Algorithms
Our lab aims to develop interpretable models of fundamental cognitive functions in primates: models that can explain basic elements of computations needed in real-world behavior. We study how the brain recognizes objects, makes decisions, and adapts its behavior to context, using a combination of neural population recordings in macaques, computational modeling, and perturbation methods.
Population Dynamics as the Language of Cognition
We approach cortical computation using the language of population geometry and dynamical systems, associating patterns of neural population activity with latent states in cognitive models. This framework is particularly powerful for cortical areas, where neurons are densely interconnected and information is encoded and transmitted across large, entangled networks of cells. In such systems, the relevant computational variables are not easily associated with individual neurons or local circuits because they emerge at the population level, in the geometry and dynamics of collective activity.
Cognitive Models as Computational Hypotheses
A central goal of the lab is to understand the brain's algorithm — the principles and equations that describe how the brain carries out its computations. We believe that cognitive models such as the drift-diffusion model are exceptionally powerful because they are both interpretable and capable of explaining a broad range of brain computations. We aim to extend these models, test their limits, and discover new principles. To ground these investigations, we primarily use perceptual decision-making tasks with naturalistic object stimuli: simple enough for rigorous quantitative modeling, yet rich enough to engage the broad range of computational elements such as generalization, flexibility, identification, and abstraction.
Perturbing Computation to Test and Treat
If we really understand mechanisms, we should be able to manipulate them. We are developing methods to perturb brain function at the level of computation. This will provide causal tests of our cognitive models and may also become potential therapeutic strategies. Our goal is to selectively modify activity in neural population state space, targeting the dynamical structures that reflect ongoing computations. As one approach, we are developing closed-loop neurofeedback paradigms that train animals to modulate their own population dynamics, or to solve tasks by reshaping the associations between neural states and external events. We believe this approach will open a principled path toward both understanding and amending the computations underlying cognitive dysfunction.
Specific Research Directions
Perceptual Decision Making
Animals must decide to act based on external events and objects. We study this process by measuring behavior and recording neural activity of animals performing tasks such as categorization of face stimuli. We have found that face categorization behavior can be understood as a process of integrating sensory evidence over multiple facial features and over time (Okazawa et al., 2018, 2021). Neural recordings from the parietal cortex revealed that neurons encode decision formation on a non-linear manifold in state space (Okazawa et al., 2021, Cell). We are now mapping how decision signals in parietal and frontal areas are formed by inputs from sensory areas encoding object information, and how flexible, context-dependent decisions emerge from these circuits.
Visual Object Recognition
For successful behavior, our brain must rapidly recognize objects present in the environment. Images are extremely rich and complicated, and the brain must extract meaningful information. We have shown that neurons in mid-level visual areas encode naturalistic textures — important components of object images — and that this selectivity can be explained by responses to higher-order statistical parameters of images (Okazawa et al., 2015, 2016). We are now investigating how neural populations in visual and higher-level areas dynamically respond during processing of complex object information, including abstract conceptual categories.