Laboratory of visual and auditory neurophysiology
We rapidly and effortlessly recognize objects and faces even though we never see the same retinal image twice. We comprehend speech across speakers, and we articulate complex motions (singing, playing instruments, dancing). Understanding how the brain recognizes and processes complex information is fundamental to many fields of science, ranging from machine learning, robotics, and space exploration, to understanding and treating neurological diseases such as autism and epilepsy.
Our brains are remarkably adept at learning and generalizing from limited experience to thrive in new environments. The computational basis of this ability is a key topic in machine learning (e.g. 'deep learning'). The challenge is how to learn good representations that enable generalization (e.g. across changes in viewing angle and illumination, or to novel instances of a class). Although recent models have rapidly improved in performance and appear to capture essential properties of early stage computations, they are missing a critical part - how to generalize from complex patterns, computations that rely on higher brain areas. Unlike early stages that are organized according to the spatial layout in sensory organs (e.g the retina or cochlea), the organizational and computational rules in later stages are unclear. Standard models lack biologically-realistic high-level organization (e.g. 'bag-of-features' approach), and a critical but missing type of data is the role of local populations of neurons in higher stages. Specifically, how do local ensembles support generalization? How do factors such as population homogeneity/heterogeneity and patterns of signal and noise correlation contribute to efficient cortical processing and behavior? Standard methods such as single electrodes and neuroimaging cannot reveal how groups of spiking neurons behave, particularly at short distances and in relation to mechanisms such as inhibition and spike timing dependent plasticity. We are acquiring the data needed to bridge this gap.
Linking neuroimaging and neurological disease to neurophysiology
Many mental illnesses (autism, Parkinson's, Lewy body dementia, epilepsy) are thought to be linked to an imbalance in inhibition and dynamics in the anterior temporal lobe. Non-invasive biomarkers are needed to detect their onset earlier, to localize and understand the mechanisms, to quantify the severity, and to develop better drugs and non-invasive treatments. Neuroimaging is a useful biomarker, but it has coarse spatiotemporal resolution (~0.1 sample/mm3, ~0.1 - 1.0 Hz) and is an indirect measure based on blood flow. Electro-encephalography (EEG) also has similar challenges.
To tie coarse neuroimaging and EEG signals to spiking populations, we are using a tightly integrated combination of approaches in the same animal: functional neuroimaging, sub-dural electrode arrays, and dense multi-depth electrode arrays ('MEA', ~100 neurons/mm3, 20 kHz). This will allow us to relate neural activity across different scales and different signal types, which is essential to bridge studies across different laboratories and different species. These studies also enable us to develop non-invasive biomarkers based on cognitive tests, which have benefits in terms of specificity, cost, and the ability to quantitatively measure day-to-day changes that are masked (undetected) in daily life and standard questionnaires.