In this paper, we study communication-efficient decentralized training of large-scale machine learning models over a network. We propose and analyze SQuARM-SGD, a decentralized training algorithm, employing momentum and compressed communication between nodes regulated by a locally computable triggering rule. In SQuARM-SGD, each node performs a fixed number of local SGD (stochastic gradient descent) steps using Nesterov's momentum and then sends sparisified and quantized updates to its neighbors only when there is a significant change in its model parameters since the last time communication occurred. We provide convergence guarantees of our algorithm for strongly-convex and non-convex smooth objectives. We believe that ours is the first theoretical analysis for compressed decentralized SGD with momentum updates. We show that SQuARM-SGD converges at rate $\mathcal{O}\left(\frac{1}{nT}\right)$ for strongly-convex objectives, while for non-convex objectives it converges at rate $\mathcal{O}\left(\frac{1}{\sqrt{nT}}\right)$, thus matching the convergence rate of \emph{vanilla} distributed SGD in both these settings. We corroborate our theoretical understanding with experiments and compare the performance of our algorithm with the state-of-the-art, showing that without sacrificing much on the accuracy, SQuARM-SGD converges at a similar rate while saving significantly in total communicated bits.

The content of this post is a partial reproduction of a chapter from the book: "Deep Learning with PyTorch Step-by-Step: A Beginner's Guide". What do gradient descent, the learning rate, and feature scaling have in common? Every time we train a deep learning model, or any neural network for that matter, we're using gradient descent (with backpropagation). We use it to minimize a loss by updating the parameters/weights of the model. The parameter update depends on two values: a gradient and a learning rate. The learning rate gives you control of how big (or small) the updates are going to be. A bigger learning rate means bigger updates and, hopefully, a model that learns faster.

As science and engineering have become increasingly data-driven, the role of optimization has expanded to touch almost every stage of the data analysis pipeline, from the signal and data acquisition to modeling and prediction. The optimization problems encountered in practice are often nonconvex. While challenges vary from problem to problem, one common source of nonconvexity is nonlinearity in the data or measurement model. Nonlinear models often exhibit symmetries, creating complicated, nonconvex objective landscapes, with multiple equivalent solutions. Nevertheless, simple methods (e.g., gradient descent) often perform surprisingly well in practice. The goal of this survey is to highlight a class of tractable nonconvex problems, which can be understood through the lens of symmetries. These problems exhibit a characteristic geometric structure: local minimizers are symmetric copies of a single ``ground truth'' solution, while other critical points occur at balanced superpositions of symmetric copies of the ground truth, and exhibit negative curvature in directions that break the symmetry. This structure enables efficient methods to obtain global minimizers. We discuss examples of this phenomenon arising from a wide range of problems in imaging, signal processing, and data analysis. We highlight the key role of symmetry in shaping the objective landscape and discuss the different roles of rotational and discrete symmetries. This area is rich with observed phenomena and open problems; we close by highlighting directions for future research.

We analyze the training dynamics for deep linear networks using a new metric - layer imbalance - which defines the flatness of a solution. We demonstrate that different regularization methods, such as weight decay or noise data augmentation, behave in a similar way. Training has two distinct phases: 1) optimization and 2) regularization. First, during the optimization phase, the loss function monotonically decreases, and the trajectory goes toward a minima manifold. Then, during the regularization phase, the layer imbalance decreases, and the trajectory goes along the minima manifold toward a flat area. Finally, we extend the analysis for stochastic gradient descent and show that SGD works similarly to noise regularization.

In this paper, we propose a novel loss function for training Generative Adversarial Networks (GANs) aiming towards deeper theoretical understanding as well as improved performance for the underlying optimization problem. The new loss function is based on cumulant generating functions giving rise to \emph{Cumulant GAN}. Relying on a recently-derived variational formula, we show that the corresponding optimization problem is equivalent to R{\'e}nyi divergence minimization, thus offering a (partially) unified perspective of GAN losses: the R{\'e}nyi family encompasses Kullback-Leibler divergence (KLD), reverse KLD, Hellinger distance and $\chi^2$-divergence. Wasserstein GAN is also a member of the proposed cumulant GAN. In terms of stability, we rigorously prove the exponential convergence of cumulant GAN to the Nash equilibrium for a linear discriminator, Gaussian distributions and the standard gradient descent algorithm. Finally, we experimentally demonstrate that image generation is generally more robust relative to Wasserstein GAN and it is substantially improved in terms of inception score when weaker discriminators are considered.

The performance of a syntax-guided synthesis algorithm is highly dependent on the provision of a good syntactic template, or grammar. Provision of such a template is often left to the user to do manually, though in the absence of such a grammar, state-of-the-art solvers will provide their own default grammar, which is dependent on the signature of the target program to be sythesized. In this work, we speculate this default grammar could be improved upon substantially. We build sets of rules, or metagrammars, for constructing grammars, and perform a gradient descent over these metagrammars aiming to find a metagrammar which solves more benchmarks and on average faster. We show the resulting metagrammar enables CVC4 to solve 26% more benchmarks than the default grammar within a 300s time-out, and that metagrammars learnt from tens of benchmarks generalize to performance on 100s of benchmarks.

We study the implicit regularization phenomenon induced by simple optimization algorithms in over-parameterized nonlinear statistical models. Specifically, we study both vector and matrix single index models where the link function is nonlinear and unknown, the signal parameter is either a sparse vector or a low-rank symmetric matrix, and the response variable can be heavy-tailed. To gain a better understanding the role of implicit regularization in the nonlinear models without excess technicality, we assume that the distribution of the covariates is known as a priori. For both the vector and matrix settings, we construct an over-parameterized least-squares loss function by employing the score function transform and a robust truncation step designed specifically for heavy-tailed data. We propose to estimate the true parameter by applying regularization-free gradient descent to the loss function. When the initialization is close to the origin and the stepsize is sufficiently small, we prove that the obtained solution achieves minimax optimal statistical rates of convergence in both the vector and matrix cases. In particular, for the vector single index model with Gaussian covariates, our proposed estimator is shown to enjoy the oracle statistical rate. Our results capture the implicit regularization phenomenon in over-parameterized nonlinear and noisy statistical models with possibly heavy-tailed data.

We study the training of regularized neural networks where the regularizer can be non-smooth and non-convex. We propose a unified framework for stochastic proximal gradient descent, which we term ProxGen, that allows for arbitrary positive preconditioners and lower semi-continuous regularizers. Our framework encompasses standard stochastic proximal gradient methods without preconditioners as special cases, which have been extensively studied in various settings. Not only that, we present two important update rules beyond the well-known standard methods as a byproduct of our approach: (i) the first closed-form proximal mappings of $\ell_q$ regularization ($0 \leq q \leq 1$) for adaptive stochastic gradient methods, and (ii) a revised version of ProxQuant that fixes a caveat of the original approach for quantization-specific regularizers. We analyze the convergence of ProxGen and show that the whole family of ProxGen enjoys the same convergence rate as stochastic proximal gradient descent without preconditioners. We also empirically show the superiority of proximal methods compared to subgradient-based approaches via extensive experiments. Interestingly, our results indicate that proximal methods with non-convex regularizers are more effective than those with convex regularizers.

We introduce in this work a new method for online approximate Bayesian learning, whose main idea is to approximate the sequence $(\pi_t)_{t\geq 1}$ of posterior distributions by a sequence $(\tilde{\pi}_t)_{t\geq 1}$ which (i) can be estimated in an online fashion using sequential Monte Carlo methods and (ii) is shown to converge to the same distribution as the sequence $(\pi_t)_{t\geq 1}$, under weak assumptions on the statistical model at hand. In its simplest version, the proposed approach amounts to take for $(\tilde{\pi}_t)_{t\geq 1}$ the sequence of filtering distributions associated to a particular state-space model, and to approximate this sequence using a standard particle filter algorithm. We illustrate on several challenging examples the benefits of this procedure for online approximate Bayesian parameter inference, and with one real data example we show that its online predictive performance can significantly outperform that of stochastic gradient descent and of streaming variational Bayes.

From the previous article, we learnt how a single neuron or perceptron works by taking the dot product of input vectors and weights,adding bias and then applying non-linear activation function to produce output.Now let's take that information and see how these neurons build up to a neural network. Now z W0 xj*wj denotes the dot product of input vectors and weights and our final output y is just activation function applied on z. Now,if we want a multi output neural network(from the diagram above),we can simply add one of these perceptrons & we have two outputs with a different set of weights and inputs.Since all the inputs are densely connected to all the outputs,these layers are also called as Dense layers.To implement this layer, we can use many libraries such keras,tensorflow,pytorch,etc. Here it shows the tensorflow implementation of this 2 perceptron network where units 2 indicate we have two outputs in this layer.We can customize this layer by adding activation function,bias constraint etc. Now,let's take a step further and let's understand how a single layer neural network works where we have a single hidden layer which feeds into the output layer. We call this a hidden layer because unlike our input and output layer which we can see or observe them.Our hidden layers are not directly observable,we can probe inside the network and see them using tools such as Netron but we can't enforce it as these are learned .