Raghav Somani

Raghav Somani

5th year Ph.D. student at
Paul G. Allen School of Computer Science and Engineering

Posts by Collection

Misc

Projects

Sparse Regression and Support Recovery bounds for Orthogonal Matching Pursuit

Posted on:

We study the problem of sparse regression where the goal is to learn a sparse vector that best optimizes a given objective function. Under the assumption that the objective function satisfies restricted strong convexity (RSC), we analyze Orthogonal Matching Pursuit (OMP) and obtain support recovery result as well as a tight generalization error bound for OMP. Furthermore, we obtain lower bounds for OMP, showing that both our results on support recovery and generalization error are tight up to logarithmic factors. To the best of our knowledge, these support recovery and generalization bounds are the first such matching upper and lower bounds (up to logarithmic factors) for any sparse regression algorithm under the RSC assumption. Read more

Universality Patterns in the Training of Neural Networks

Posted on:

This work proposes and demonstrates a surprising pattern in the training of neural networks: there is a one to one relation between the values of any pair of losses (such as cross entropy, mean squared error, \(0/1\) error etc.) evaluated for a model arising at (any point of) a training run. This pattern is universal in the sense that this one to one relationship is identical across architectures (such as VGG, Resnet, Densenet etc.), algorithms (SGD and SGD with momentum) and training loss functions (cross entropy and mean squared error). Read more

Scaling laws of optimization algorithms for Deep Learning - the Graphon perspective

Posted on:

Stochastic optimization algorithms are used to train large (Deep) Neural Networks ((D)NNs). The non-linear training dynamics of two-layer NNs can be described as a mean-field interacting particle system where the “particles” are the neurons in the single hidden layer. Wasserstein gradient flows often arise from such mean-field interactions among exchangeable particles. This relies on the observation that the permutation symmetry of the neurons in the hidden layer allows the problem to be viewed as an optimization problem over probability measures. Going beyond, multi-layer NNs can be considered as large computational graphs and therefore can possess different groups of symmetries. This body of work aims to describe analytical scaling limits of stochastic optimization algorithms as the size of the network grows. We use and develop on the existing theory of exchangeable arrays, graphons (analytical limits of dense graphs), the general theory of gradient flows on metric spaces, and insights from propagation of chaos to characterize this scaling limit. We discover a generalized notion of the McKean-Vlasov equation on graphons where the phenomenon of propagation of chaos holds. In the asymptotically zero-noise case, this limit is a gradient flow on the metric space of graphons. Read more

Publications

Non-Gaussianity of Stochastic Gradient Noise

Abhishek Panigrahi, Raghav Somani, Navin Goyal & Praneeth Netrapalli
Published at: Science meets Engineering of Deep Learning (SEDL) workshop, Neural Information Processing Systems (NeurIPS), 2019

We study the distribution of the Stochastic Gradient Noise during the training and observe that for batch sizes \(256\) and above, the distribution is best described as Gaussian at-least in the early phases of training. Read more

[arXiv] [bib]