webinar register page

Webinar banner
Amy Kuceyeski: Biological and Artificial Neural Networks
The recent explosion of machine learning literature has centered largely around Artificial Neural Networks (ANNs). These networks, originally inspired by biological neural networks - specifically, how the human brain processes visual information (Rosenblatt et al., 1958) - have proved remarkably useful for classification or regression problems of many types. Meanwhile, in the field of neuroscience, researchers have incorporated ANNs into "encoding models" that predict neural responses to visual stimuli and, furthermore, have been shown to reflect structure and function of the visual processing pathway. This observation has led to speculation that primate ventral visual stream may have evolved to be an optimal system for object recognition/detection in the same way that ANNs are identifying optimal computational architectures. Here, we introduce NeuroGen, a novel encoding/generative model architecture designed to synthesize realistic images predicted to maximize or minimize activation in pre-selected regions of the human visual cortex. We then apply this framework as a discovery architecture to amplify differences in regional and individual brain response patterns to visual stimuli, and, furthermore, use it to generate synthetic images predicted to achieve levels of activation above and beyond what is achievable with natural images. If it can be shown with future work that the synthetic images actually produce the desired target responses, this approach could be used to perform macro-scale, non-invasive neuronal population control in humans.
Webinar logo
* Required information
Loading