Why do we study the brain?
We are interested in the brain because the brain supports a remarkable set of behavior. These behaviors define who we are. And it is very difficult to live without them. We want to “understand” how does the brain support these behaviors. We are especially interested in studying visually driven behavior like dynamic scene understanding.
Operationalizing the behavior of choice
So first things first — we need to coarsely identify the behavior of choice. But saying “attention”, “memory”, “object perception” etc. won’t do. To figure out whether our “understanding” is real, we will need to operationalize these behaviors in terms of objective behavioral tasks where human performance at different granularities can be measured by doing experiments. The next goal is to conduct behavioral experiments to measure this behavior in humans (we are ultimately interested in human visual cognition). Once we have defined the behavior and measured it — we are in good shape!
Choice of animal model
Choice of animal model: To successfully probe the neural mechanisms of human visual intelligence, we need an appropriate animal model that meets three essential criteria. First, to interchangeably test a variety of stimuli and tasks between the animal model and human subjects, the animals should achieve human-like performance in visuocognitive tasks. Second, it should be possible to conduct fine-grain (single neuron level) measurements and targeted causal perturbations to interrogate neural circuits and test specific hypotheses. Third, inferences made on the animals’ neural mechanisms should be relevant for humans due to established similarities in brain areas and behavior. The rhesus macaque (with its ventral visual cortex that supports human-like object and face processing) is a remarkable example that meets all these criteria – and therefore is our animal model of choice.
Human crowd-sourced psychophysics
How does the brain support the specific behaviors?
What brain areas are causally involved in the behavior? How are the visual signals transformed in the brain and reach a form that is consistent with the measured behavior? One level of “understanding” for us will arise if we can predict the human/NHP behavior (at varied grains of measurements) from the dynamic brain activity (combined optimally across different areas; where that combination will serve as the leading decoding hypothesis). Our methods (e.g. NHP electrophysiology, in-vivo genetic technology-based probes, etc.) are all directed towards establishing those links.
In-vivo genetic approaches
Developing and utilizing models of the brain
Once we have coarsely identified the implicated brain regions, we can start building detailed encoding models of the responses in those areas. So another form of “understanding” we seek is detailed neuro-mechanistic models of the implicated brain areas (with the explicit prediction of the measured signal like spike rate) that are indistinguishable from the brain data. As it stands now, the best encoding models are being generated by folks in Artificial Intelligence (AI). We are cognizant enough that we acknowledge that and we will build on that line of work. But AI and Neuroscience have different goals. While it is likely that two systems (AI and human brain) that perform the same tasks will have some level of shared internal mechanisms, it is our responsibility as neuroscientists to not give in to the AI hype and treat all models from ML as achievements in neuroscience. Currently, in the absence of serious (non-toy), predictive modeling from computational neuroscience, the models coming out of computer vision are our best bets (starting hypotheses). But the ViTA lab aims to put these models to strong tests and as well measure neural data that provides significant guard rails to transform these AI models into better models of the brain. While all this fun modeling endeavors are being pursued — one might ask: What do you get out of building these complex models to seek understanding of a system? It sure doesn’t meet the interpretability criteria for many — so not good for an intuitive understanding… We seek to build these detailed computational models so that we can use them to reverse engineer solutions for neurological disorders.