Although distributed computing can significantly reduce the training time of deep neural networks, scaling the training process while maintaining high efficiency and final accuracy is challenging. Distributed asynchronous training enjoys near-linear speedup, but asynchrony causes gradient staleness, the main difficulty in scaling stochastic gradient descent to large clusters. Momentum, which is often used to accelerate convergence and escape local minima, exacerbates the gradient staleness, thereby hindering convergence. We propose DANA: a novel asynchronous distributed technique which is based on a new gradient staleness measure that we call the gap. By minimizing the gap, DANA mitigates the gradient staleness, despite using momentum, and therefore scales to large clusters while maintaining high final accuracy and fast convergence. DANA adapts Nesterov's Accelerated Gradient to a distributed setting, computing the gradient on an estimated future position of the model's parameters. In turn, we show that DANA's estimation of the future position amplifies the use of a Taylor expansion, which relies on a fast Hessian approximation, making it much more effective and accurate. Our evaluation on the CIFAR and ImageNet datasets shows that DANA outperforms existing methods, in both final accuracy and convergence speed.

Semi-supervised learning (SSL) plays an increasingly important role in the big data era because a large number of unlabeled samples can be used effectively to improve the performance of the classifier. Semi-supervised support vector machine (S$^3$VM) is one of the most appealing methods for SSL, but scaling up S$^3$VM for kernel learning is still an open problem. Recently, a doubly stochastic gradient (DSG) algorithm has been proposed to achieve efficient and scalable training for kernel methods. However, the algorithm and theoretical analysis of DSG are developed based on the convexity assumption which makes them incompetent for non-convex problems such as S$^3$VM. To address this problem, in this paper, we propose a triply stochastic gradient algorithm for S$^3$VM, called TSGS$^3$VM. Specifically, to handle two types of data instances involved in S$^3$VM, TSGS$^3$VM samples a labeled instance and an unlabeled instance as well with the random features in each iteration to compute a triply stochastic gradient. We use the approximated gradient to update the solution. More importantly, we establish new theoretic analysis for TSGS$^3$VM which guarantees that TSGS$^3$VM can converge to a stationary point. Extensive experimental results on a variety of datasets demonstrate that TSGS$^3$VM is much more efficient and scalable than existing S$^3$VM algorithms.

Minimizing a convex function of a measure with a sparsity-inducing penalty is a typical problem arising, e.g., in sparse spikes deconvolution or two-layer neural networks training. We show that this problem can be solved by discretizing the measure and running non-convex gradient descent on the positions and weights of the particles. For measures on a $d$-dimensional manifold and under some non-degeneracy assumptions, this leads to a global optimization algorithm with a complexity scaling as $\log(1/\epsilon)$ in the desired accuracy $\epsilon$, instead of $\epsilon^{-d}$ for convex methods. The key theoretical tools are a local convergence analysis in Wasserstein space and an analysis of a perturbed mirror descent in the space of measures. Our bounds involve quantities that are exponential in $d$ which is unavoidable under our assumptions.

While stochastic gradient descent (SGD) and variants have been surprisingly successful for training deep nets, several aspects of the optimization dynamics and generalization are still not well understood. In this paper, we present new empirical observations and theoretical results on both the optimization dynamics and generalization behavior of SGD for deep nets based on the Hessian of the training loss and associated quantities. We consider three specific research questions: (1) what is the relationship between the Hessian of the loss and the second moment of stochastic gradients (SGs)? (2) how can we characterize the stochastic optimization dynamics of SGD with fixed and adaptive step sizes and diagonal pre-conditioning based on the first and second moments of SGs? and (3) how can we characterize a scale-invariant generalization bound of deep nets based on the Hessian of the loss, which by itself is not scale invariant? We shed light on these three questions using theoretical results supported by extensive empirical observations, with experiments on synthetic data, MNIST, and CIFAR-10, with different batch sizes, and with different difficulty levels by synthetically adding random labels.

We consider a decentralized learning problem, where a set of computing nodes aim at solving a non-convex optimization problem collaboratively. It is well-known that decentralized optimization schemes face two major system bottlenecks: stragglers' delay and communication overhead. In this paper, we tackle these bottlenecks by proposing a novel decentralized and gradient-based optimization algorithm named as QuanTimed-DSGD. Our algorithm stands on two main ideas: (i) we impose a deadline on the local gradient computations of each node at each iteration of the algorithm, and (ii) the nodes exchange quantized versions of their local models. The first idea robustifies to straggling nodes and the second alleviates communication efficiency. The key technical contribution of our work is to prove that with non-vanishing noises for quantization and stochastic gradients, the proposed method exactly converges to the global optimal for convex loss functions, and finds a first-order stationary point in non-convex scenarios. Our numerical evaluations of the QuanTimed-DSGD on training benchmark datasets, MNIST and CIFAR-10, demonstrate speedups of up to 3x in run-time, compared to state-of-the-art decentralized optimization methods.

Nonconvex optimization algorithms with random initialization have attracted increasing attention recently. It has been showed that many first-order methods always avoid saddle points with random starting points. In this paper, we answer a question: can the nonconvex heavy-ball algorithms with random initialization avoid saddle points? The answer is yes! Direct using the existing proof technique for the heavy-ball algorithms is hard due to that each iteration of the heavy-ball algorithm consists of current and last points. It is impossible to formulate the algorithms as iteration like xk+1= g(xk) under some mapping g. To this end, we design a new mapping on a new space. With some transfers, the heavy-ball algorithm can be interpreted as iterations after this mapping. Theoretically, we prove that heavy-ball gradient descent enjoys larger stepsize than the gradient descent to escape saddle points to escape the saddle point. And the heavy-ball proximal point algorithm is also considered; we also proved that the algorithm can always escape the saddle point.

We study distributed algorithms for expected loss minimization where the datasets are large and have to be stored on different machines. Often we deal with minimizing the average of a set of convex functions where each function is the empirical risk of the corresponding part of the data. In the distributed setting where the individual data instances can be accessed only on the local machines, there would be a series of rounds of local computations followed by some communication among the machines. Since the cost of the communication is usually higher than the local machine computations, it is important to reduce it as much as possible. However, we should not allow this to make the computation too expensive to become a burden in practice. Using second-order methods could make the algorithms converge faster and decrease the amount of communication needed. There are some successful attempts in developing distributed second-order methods. Although these methods have shown fast convergence, their local computation is expensive and could enjoy more improvement for practical uses. In this study we modify an existing approach, DANE (Distributed Approximate NEwton), in order to improve the computational cost while maintaining the accuracy. We tackle this problem by using iterative methods for solving the local subproblems approximately instead of providing exact solutions for each round of communication. We study how using different iterative methods affect the behavior of the algorithm and try to provide an appropriate tradeoff between the amount of local computation and the required amount of communication. We demonstrate the practicality of our algorithm and compare it to the existing distributed gradient based methods such as SGD.

We investigate a sequential optimization procedure to minimize the empirical risk functional $f_{\hat\theta}(x) = \frac{1}{2}\|G_{\hat\theta}(x) - y\|^2$ for certain families of deep networks $G_{\theta}(x)$. The approach is to optimize a sequence of objective functions that use network parameters obtained during different stages of the training process. When initialized with random parameters $\theta_0$, we show that the objective $f_{\theta_0}(x)$ is "nice'' and easy to optimize with gradient descent. As learning is carried out, we obtain a sequence of generative networks $x \mapsto G_{\theta_t}(x)$ and associated risk functions $f_{\theta_t}(x)$, where $t$ indicates a stage of stochastic gradient descent during training. Since the parameters of the network do not change by very much in each step, the surface evolves slowly and can be incrementally optimized. The algorithm is formalized and analyzed for a family of expansive networks. We call the procedure {\it surfing} since it rides along the peak of the evolving (negative) empirical risk function, starting from a smooth surface at the beginning of learning and ending with a wavy nonconvex surface after learning is complete. Experiments show how surfing can be used to find the global optimum and for compressed sensing even when direct gradient descent on the final learned network fails.

The mushroom body is the key network for the representation of learned olfactory stimuli in Drosophila and insects. The sparse activity of Kenyon cells, the principal neurons in the mushroom body, plays a key role in the learned classification of different odours. In the specific case of the fruit fly, the sparseness of the network is enforced by an inhibitory feedback neuron called APL, and by an intrinsic high firing threshold of the Kenyon cells. In this work we took inspiration from the fruit fly brain to formulate a novel machine learning algorithm that is able to optimize the sparsity level of a reservoir by changing the firing thresholds of the nodes. The sparsity is only applied on the readout layer so as not to change the timescales of the reservoir and to allow the derivation of a one-layer update rule for the firing thresholds. The proposed algorithm is a combination of learning a neuron-specific sparsity threshold via gradient descent and a global sparsity threshold via a Markov chain Monte Carlo method. The proposed model outperforms the standard gradient descent, which is limited to the readout weights of the reservoir, on two example tasks. It demonstrates how the learnt sparse representation can lead to better classification performance, memorization ability and convergence time.

Rectified Linear Units (ReLU) have become the main model for the neural units in current deep learning systems. This choice has been originally suggested as a way to compensate for the so called vanishing gradient problem which can undercut stochastic gradient descent (SGD) learning in networks composed of multiple layers. Here we provide analytical results on the effects of ReLUs on the capacity and on the geometrical landscape of the solution space in two-layer neural networks with either binary or real-valued weights. We study the problem of storing an extensive number of random patterns and find that, quite unexpectedly, the capacity of the network remains finite as the number of neurons in the hidden layer increases, at odds with the case of threshold units in which the capacity diverges. Possibly more important, a large deviation approach allows us to find that the geometrical landscape of the solution space has a peculiar structure: while the majority of solutions are close in distance but still isolated, there exist rare regions of solutions which are much more dense than the similar ones in the case of threshold units. These solutions are robust to perturbations of the weights and can tolerate large perturbations of the inputs. The analytical results are corroborated by numerical findings.