My research interest originates from the desire to understand the visual representation in the (primate) brains and machines. In our visual brain, the pattern of photons hitting our retina is transformed by a series of visual cortices and represented by different neuron firing patterns. These intermediate neural representations bear no resemblance to the image hitting retina, but they support complex perceptions like object recognition. The neural code of images is of primary interest in my Ph.D work. 

To a first approximation, the mean responses of a visual neuron is a continuous function of the retinal input. Conceptually, we can think about each neuron as defining a landscape over all the whole image manifold: the peaks are domains of images eliciting high response and plains are domains of images eliciting low to no responses. In this picture, we could redefine the questions of visual neuroscience. Where do the neurons place their peaks? How do the activation change when deviating from the peaks? How the tuning landscape supports invariant object code? The dream is that through the lens of geometry, we could uncover obscured relationship in neural tuning. 

This research interest branches into a few different directions. 

What is the geometry of the natural image manifold? To understand the neural tuning on the natural image manifold, we'd better understand the structure of this manifold first. To make this question concrete, we analyzed the geometry of image manifolds parametrized by Generative Adversarial Networks (GAN). The result became widely useful in accelerating search on image manifold, and analyzing neural tuning on the manifold. 

How to search for peaks on the tuning landscape efficiently? To understand the tuning landscape, one obvious strategy is to start from the peaks, since they are rare and define the property of the neuron. But to search for these peaks requires black box optimization algorithms optimizing a noisy neuronal response. With large scale in silico benchmarking, we found that CMAES excelled in its ability to optimize noisy neuronal responses. Using knowledge about the structure of image manifold, we could further accelerate the search algorithms. 

How to characterize the shape of tuning landscape?

How to model the neural tuning using the Evolution trajectory? 

How visual representation came about through unsupervised learning? At birth, animals do not have semantic labels directly available to their visual cortices. Thus, the representation in visual brain has to be learned by some ways other than supervised learning. In this project, I tried to provide some initial evidence that, using naturalistic viewing behaviors (saccades), and constrastive loss (SimCLR), the visual cortices could learn a representation that support linear object classification.