This is a nice talk that discusses, among other things, subleading 1/width corrections to the infinite width limit of neural networks. I was expecting someone would work out these corrections when I wrote the post on NTK and large width limit at the link below. Apparently, the infinite width limit does not capture the behavior of realistic neural nets and it is only at the first nontrivial order in the expansion that the desired properties emerge. Roberts claims that when the depth to width ratio r is small but nonzero one can characterize network dynamics in a controlled expansion, whereas when r > 1 it becomes a problem of strong dynamics.
The talk is based on the book
This book develops an effective theory approach to understanding deep neural networks of practical relevance. Beginning from a first-principles component-level picture of networks, we explain how to determine an accurate description of the output of trained networks by solving layer-to-layer iteration equations and nonlinear learning dynamics. A main result is that the predictions of networks are described by nearly-Gaussian distributions, with the depth-to-width aspect ratio of the network controlling the deviations from the infinite-width Gaussian description. We explain how these effectively-deep networks learn nontrivial representations from training and more broadly analyze the mechanism of representation learning for nonlinear models. From a nearly-kernel-methods perspective, we find that the dependence of such models' predictions on the underlying learning algorithm can be expressed in a simple and universal way. To obtain these results, we develop the notion of representation group flow (RG flow) to characterize the propagation of signals through the network. By tuning networks to criticality, we give a practical solution to the exploding and vanishing gradient problem. We further explain how RG flow leads to near-universal behavior and lets us categorize networks built from different activation functions into universality classes. Altogether, we show that the depth-to-width ratio governs the effective model complexity of the ensemble of trained networks. By using information-theoretic techniques, we estimate the optimal aspect ratio at which we expect the network to be practically most useful and show how residual connections can be used to push this scale to arbitrary depths. With these tools, we can learn in detail about the inductive bias of architectures, hyperparameters, and optimizers.
Dan Roberts web page.
This essay looks interesting:
We discuss why AI is hard and why physics is simple. We discuss how physical intuition and the approach of theoretical physics can be brought to bear on the field of artificial intelligence and specifically machine learning. We suggest that the underlying project of machine learning and the underlying project of physics are strongly coupled through the principle of sparsity, and we call upon theoretical physicists to work on AI as physicists. As a first step in that direction, we discuss an upcoming book on the principles of deep learning theory that attempts to realize this approach.
May 2021 post: Neural Tangent Kernels and Theoretical Foundations of Deep LearningSee related remarks: ICML notes (2018).
Large width seems to provide a limiting case (analogous to the large-N limit in gauge theory) in which rigorous results about deep learning can be proved. ...
The overparametrized (width ~ w^2) network starts in a random state and by concentration of measure this initial kernel K is just the expectation, which is the NTK. Because of the large number of parameters the effect of training (i.e., gradient descent) on any individual parameter is 1/w, and the change in the eigenvalue spectrum of K is also 1/w. It can be shown that the eigenvalue spectrum is positive and bounded away from zero, and this property does not change under training. Also, the evolution of f is linear in K up to corrections with are suppressed by 1/w. Hence evolution follows a convex trajectory and can achieve global minimum loss in a finite (polynomial) time.
The parametric 1/w expansion may depend on quantities such as the smallest NTK eigenvalue k: the proof might require k >> 1/w or wk large.
In the large w limit the function space has such high dimensionality that any typical initial f is close (within a ball of radius 1/w?) to an optimal f. These properties depend on specific choice of loss function.
It may turn out that the problems on which DL works well are precisely those in which the training data (and underlying generative processes) have a hierarchical structure which is sparse, level by level. Layered networks perform a kind of coarse graining (renormalization group flow): first layers filter by feature, subsequent layers by combinations of features, etc. But the whole thing can be understood as products of sparse filters, and the performance under training is described by sparse performance guarantees (ReLU = thresholded penalization?). Given the inherent locality of physics (atoms, molecules, cells, tissue; atoms, words, sentences, ...) it is not surprising that natural phenomena generate data with this kind of hierarchical structure.