The Active Vision and Neural Computation Lab
Research in the lab is broadly focused on how primate brains extract information about the world using vision during natural behavior.
We want better models of the visual system because good models of vision are useful. Early vision forms the fundamental building blocks of our conscious visual perception and has been a useful testbed for theories of neural coding. Good models of early vision are useful for scientists studying cognition in natural tasks and environments. Moreover, being able to generalize and predict how neurons at different levels in the visual pathway would respond to a particular input is useful for treating human diseases and driving technological development. For example, what information is lost, specifically, when a particular cell class in the retina is damaged? What parameters for a retinal prosthesis matter to drive visual cortex normally? What parameters matter to avoid artifacts in the development of new displays? How can we make computer vision more robust to adversarial examples? All of these questions are answerable if we had a good-enough model of the early visual system. To that end, models we develop compute based on realistic input (i.e., 2D spatiotemporal luminance sequences) and produce predictions for neural activity at multiple stages of the visual pathway, including perception.
Humans are primates. Like other primates, we have forward-facing eyes and a large portion of our brains devoted to processing visual input – The single largest area of your cortex is primary visual cortex (V1), the second largest cortical brain area is V2! This is true of all primates. Primates are the only mammals who have a fovea, which is a pit in the retina with tightly-packed photoreceptors that lets us see with high acuity. And we move our eyes to position the fovea over visual details of interest. One fundamental premise of our philosophy is that each species’ visual system co-evolved with its motor system. Therefore, even a purely feed-forward visual system will have learned to operate on a set of spatiotemporal statistics that are species-specific. In other words, the system expects certain types of input and is designed process those best. Thus, if we want to make good models of human vision, we need to study systems that work similarly to humans. We do want to understand human vision because our research should be useful for those who aim to treat human patients, those who design computer vision systems or develop new display technology… for humans.
Approach We approach our research questions by collaborating closely with neurophysiologists. We help design experiments and analyze neural data collected in other labs using statistical models and machine learning. We also develop new eye-tracking technology in-house and test perceptual predictions from our neural models using psychophysics with high-resolution eye-tracking.
Open positions and pre-requisites
I am looking for students and postdocs to join my lab. The work we do is fairly technical, so it is helpful to have some existing skills in machine learning and statistical modeling.
Students and postdocs will be encouraged to connect across these multiple topics.
Postdoc/PhD student for Machine Learning Models of Visual Neural Activity
I am recruiting a highly motivated and skilled postdoc and PhD student to develop and analyze predictive machine learning models for cortical and subcortical visual areas in free-viewing monkeys. Collaborations are ongoign with Professors Jude Mitchell, Alex Huk, Bevil Conway, and Greg Horwitz. For more information on the data and approach, see our recent preprint.
The candidate is expected to develop novel data-driven CNN-based machine learning models for recorded neurons in LGN, V1, V2, V4, and MT cortex, analyze the trained models to better understand how populations of neurons in visual cortex are modulated by behavior and eye movements. The cadidate will develop predictions that can be tested in subsequent neurophysiological or psychophysical experiments. The position is purely computational, but the candidate is expected to work closely with the experimental collaborators.
The ideal candidate has a degree in machine learning, physics, math, electrical engineering, or related fields, and a strong background in mathematics, machine learning, or statistics, with prior experience in (computational) neuroscience.
How to apply:
If you’re interested in joining the lab, I’d love to hear from you! To make sure it’s a good fit for you and the lab, email me with the following information:
- Your Curriculum Vitae.
- A short statement of purpose outlining your interest in this research area.
- A short list of your existing skills and those you would like to acquire.
- Known programming languages and experience with pytorch. Any example code (e.g., on github or bitbucket).
- Possible start date.
The subject line should indicate whether you are interested in a postdoc of PhD student position.
Current research topics:
Most animals with complex spatial vision use image-forming eyes and a “saccade and fixate” pattern of eye movements to see the world. However, their eyes are never still, counter-rotating relative to body and/or head movements, and drifting during “fixations”, such that the input to the retina is better thought of as a spatiotemporal movie instead of a stable (or unstable) image. My research aims to understand the algorithms the brain uses (in cortical visual areas) to utilize information that is generated by the motion of the eyes. To approach this, I use a combination of high-resolution eye-tracking and statistical models of both the visual input and neural activity in visual cortex.
Humans see best at the very center center of their visual field. This “high-resolution” region is called the fovea and, among mammals, only primates have one. The primate fovea is a highly-specialized anatomical adaptation for high-resolution spatial vision and it differs substantially from the peripheral retina and the retinas of other mammals.