Zoom link: https://upenn.zoom.us/j/98220304722

This talk reviews a theory of the functions learned by neural networks with Rectified Linear Unit (ReLU) activations. At its core is the observation that deep ReLU networks can be characterized as solutions to data-fitting problems in certain infinite dimensional function spaces. The solutions are compositions of functions from Banach spaces of second-order bounded variation, defined in the Radon transform domain. Functions in these spaces exhibit strong smoothness in most directions, making it a natural setting for adapting to intrinsic low-dimensional structure in data. Moreover, the norms in these spaces are closely tied to the size of neural network weights, providing a direct connection between function complexity and network parametrization. In particular, the total variation norm provides an analytic tool for identifying functions that cannot be realized by shallow networks, thereby yielding a precise characterization of depth separation. Representer theorems reveal the solutions are sparse in the number of active neurons per layer. Sparsity provides a principled path to network compression, yet some sparse solutions can suffer from poor generalization. The theory suggests new training strategies to promote solutions that generalize more robustly.