Simultaneous advances in neural interfacing and data science have created a remarkable opportunity to reshape our understanding of the brain in health and disease. Unfortunately, the most powerful advances in modern machine learning sometimes provide limited scientific insight due to the opaqueness of these "black box" techniques. To illustrate, I will describe our recent efforts collecting local field potential (LFP) data from participants undergoing subcallosal cingulate cortex (SCC) deep brain stimulation (DBS) as an experimental therapy for treatment-resistant depression. While SCC DBS can be effective, scaling this technique to an approved therapy requires objective characterization of patient-specific disease trajectories toward stable recovery. I will describe our development of a novel "explainable AI" technique for generating causal post-hoc explanations of black-box classifiers based on a learned low-dimensional representation of unsupervised data. Using this data-driven approach, we identify unexpected changes in SCC LFP dynamics that correspond to individual recovery trajectories, behavioral changes (reflected in facial expressions), and white matter structural integrity. In addition to illuminating changes to neural dynamics with chronic SCC DBS, our results demonstrate potential biomarkers for future clinical trials and closed-loop neuromodulation systems.
(Georgia Institute of Technology, 2017-03-15)
Rozell, Christopher J.
The natural stimuli that biological vision must use to understand the world are extremely complex.
Recent advances in machine learning have shown that low-dimensional geometric models (e.g.,
sparsity, manifolds) can capture much of the structure in complex natural images. I will describe our
work building efficient neural coding models that optimally exploit this structure. These results
incorporate the constraints of biophysical systems and the physical world by drawing on
mathematical tools such as dynamical systems, optimization, unsupervised learning, randomized
dimensionality reduction, and manifold learning. These results show that incorporating natural
constraints can lead to theoretical models that account for a wide range of observed phenomenon,
including complex response properties of individual neurons, architectural features of the network
(e.g., makeup of different cell types), and reported perceptual results from human psychophysical
experiments.