Please tell us about your lab’s research.
We study human vision, focusing on the neural and computational basis of high-level vision using cognitive and computational neuroscience techniques. Our research involves brain imaging in healthy adults while viewing sensory stimuli to map brain processes that generate behaviorally relevant representations, and we build computational models of how neurons transform sensory inputs into high-level representations.
What are artificial neural networks?
Artificial neural networks are simplified models of how neurons function in the brain—mathematical operations that mimic what real neurons do. They can perform tasks supported by regions of the human visual system, such as recognizing objects, understanding the spatial layout of the environment, and perceiving the 3D structure of space. These types of object recognition have traditionally been challenging for computer scientists to solve.
From a neuroscience perspective, what makes these things interesting is that the artificial neural networks were inspired by the principles of the brain—how it performs these computations. These networks learn through training: you show them images, correct their mistakes, and they update their internal connections to improve performance. The result is that the network shows many similarities to what appears to be happening in human vision. They also become powerful tools for neuroscientists to study, because we now have a model of how vision might be implemented in the brain.
Who are your collaborators at Johns Hopkins?
My primary collaborators are Leyla Isik in cognitive science at Hopkins, who works on high-level human vision and neural network modeling, and Brice Ménard in the physics department. We share an interest in the computational and mathematical properties of vision. Most of our current work focuses on the statistical properties of neural representations, so we study actual recordings from human fMRI and mouse calcium imaging. We’re interested in questions that address the characterization of neural representations and high-dimensional properties.
The idea is that a population of neurons in a brain region could number thousands or millions, supporting tasks associated with that region. Neuroscientists try to understand what all these neurons are doing to perform this task. One way neuroscientists approach this is to find whether there is some lower-dimensional view of this very high-dimensional system that can tell us something about its internal representation. They ask: What are the key properties of the representations that allow it to perform whatever its computational goal is?
From a theoretical perspective, it has been argued that the brain fundamentally runs in a low-dimensional way. So, even though there might be many neurons in a brain system, those neurons might all be acting in a coordinated way, with what is going on described by a relatively small number of latent factors. However,
prior data sets have usually been small, and experiments have been relatively simple, possibly biasing our view of what is happening in the brain. With larger, more complex datasets, a much higher-dimensional perspective emerges. We have been trying to characterize this high-dimensional structure of neural representations.
Aside from science, what other interests, hobbies, or passions do you have?
I love biking. I commute to campus by bike and ride my three-year-old son to school every day on a cargo bike. He has his little canopy in the back. I also enjoy mountain biking.
I love hiking. My wife and I enjoy going to Shenandoah National Park nearby and also to the national parks in Colorado, Montana, and California.
Another passion of mine is cooking, especially pizza making. I have a pizza oven in my yard, and I make my own pizza dough and cook my own pizzas.