How do neural networks learn? What do they learn? How do they represent the world they see?
Neural nets can perform complex tasks but what and how they learn is not understood. As Artificial Intelligence and Machine Learning make rapid strides, physicists at JHU are working to understand these systems and incorporate them into Physics and Astronomy research. With their large numbers of neurons and connections, neural nets can be analyzed through the lens of statistical mechanics. We are working to characterize feature learning, optimization, and the scaling of ML systems using various datasets: images, text, the web, scientific measurements, etc.