We present a direct (primal only) derivation of Mirror Descent as a "partial" discretization of gradient flow on a Riemannian manifold where the metric tensor is the Hessian of the Mirror Descent potential function. We argue that this discretization is more faithful to the geometry than Natural Gradient Descent, which is obtained by a "full" forward Euler discretization. This view helps shed light on the relationship between the methods and allows generalizing Mirror Descent to any Riemannian geometry, even when the metric tensor is not a Hessian, and thus there is no "dual".

We propose to optimize neural networks with a uniformly-distributed random learning rate. The associated stochastic gradient descent algorithm can be approximated by continuous stochastic equations and analyzed with the Fokker-Planck formalism. In the small learning rate approximation, the training process is characterized by an effective temperature which depends on the average learning rate, the mini-batch size and the momentum of the optimization algorithm. By comparing the random learning rate protocol with cyclic and constant protocols, we suggest that the random choice is generically the best strategy in the small learning rate regime, yielding better regularization without extra computational cost. We provide supporting evidence through experiments on both shallow, fully-connected and deep, convolutional neural networks for image classification on the MNIST and CIFAR10 datasets.

Rapid advances in data collection and processing capabilities have allowed for the use of increasingly complex models that give rise to nonconvex optimization problems. These formulations, however, can be arbitrarily difficult to solve in general, in the sense that even simply verifying that a given point is a local minimum can be NPhard [1]. Still, some relatively simple algorithms have been shown to lead to surprisingly good empirical results in many contexts of interest. Perhaps the most prominent example is the success of the backpropagation algorithm for training neural networks. Several recent works have pursued rigorous analytical justification for this phenomenon by studying the structure of the nonconvex optimization problems and establishing that simple algorithms, such as gradient descent and its variations, perform well in converging towards local minima and avoiding saddle-points. A key insight in these analyses is that gradient perturbations play a critical role in allowing local descent algorithms to efficiently distinguish desirable from undesirable stationary points and escape from the latter. In this article, we cover recent results on second-order guarantees for stochastic first-order optimization algorithms in centralized, federated, and decentralized architectures. A key desirable feature of automated learning algorithms is the ability to learn models directly from data with minimal need for direct intervention by the designer. The authors are with the Institute of Electrical Engineering, École Polytechnique Fédérale de Lausanne.

In this paper we analyze the necessary number of samples to estimate the gradient of any multidimensional smooth (possibly non-convex) function in a zero-order stochastic oracle model. In this model, an estimator has access to noisy values of the function, in order to produce the estimate of the gradient. We also provide an analysis on the sufficient number of samples for the finite difference method, a classical technique in numerical linear algebra. For $T$ samples and $d$ dimensions, our information-theoretic lower bound is $\Omega(\sqrt{d/T})$. We show that the finite difference method has rate $O(d^{4/3}/\sqrt{T})$ for functions with zero third and higher order derivatives. Thus, the finite difference method is not minimax optimal, and therefore there is space for the development of better gradient estimation methods.

We propose a novel regularization-based continual learning method, dubbed as Adaptive Group Sparsity based Continual Learning (AGS-CL), using two group sparsity-based penalties. Our method selectively employs the two penalties when learning each node based its the importance, which is adaptively updated after learning each new task. By utilizing the proximal gradient descent method for learning, the exact sparsity and freezing of the model is guaranteed, and thus, the learner can explicitly control the model capacity as the learning continues. Furthermore, as a critical detail, we re-initialize the weights associated with unimportant nodes after learning each task in order to prevent the negative transfer that causes the catastrophic forgetting and facilitate efficient learning of new tasks. Throughout the extensive experimental results, we show that our AGS-CL uses much less additional memory space for storing the regularization parameters, and it significantly outperforms several state-of-the-art baselines on representative continual learning benchmarks for both supervised and reinforcement learning tasks.

This paper considers the problem of multi-agent distributed optimization. In this problem, there are multiple agents in the system, and each agent only knows its local cost function. The objective for the agents is to collectively compute a common minimum of the aggregate of all their local cost functions. In principle, this problem is solvable using a distributed variant of the traditional gradient-descent method, which is an iterative method. However, the speed of convergence of the traditional gradient-descent method is highly influenced by the conditioning of the optimization problem being solved. Specifically, the method requires a large number of iterations to converge to a solution if the optimization problem is ill-conditioned. In this paper, we propose an iterative pre-conditioning approach that can significantly attenuate the influence of the problem's conditioning on the convergence-speed of the gradient-descent method. The proposed pre-conditioning approach can be easily implemented in distributed systems and has minimal computation and communication overhead. For now, we only consider a specific distributed optimization problem wherein the individual local cost functions of the agents are quadratic. Besides the theoretical guarantees, the improved convergence speed of our approach is demonstrated through experiments on a real data-set.

With the proliferation of smart devices having built-in sensors, Internet connectivity, and programmable computation capability in the era of Internet of things (IoT), tremendous data is being generated at the network edge. Federated learning is capable of analyzing the large amount of data from a distributed set of smart devices without requiring them to upload their data to a central place. However, the commonly-used federated learning algorithm is based on stochastic gradient descent (SGD) and not suitable for resource-constrained IoT environments due to its high communication resource requirement. Moreover, the privacy of sensitive data on smart devices has become a key concern and needs to be protected rigorously. This paper proposes a novel federated learning framework called DP-PASGD for training a machine learning model efficiently from the data stored across resource-constrained smart devices in IoT while guaranteeing differential privacy. The optimal schematic design of DP-PASGD that maximizes the learning performance while satisfying the limits on resource cost and privacy loss is formulated as an optimization problem, and an approximate solution method based on the convergence analysis of DP-PASGD is developed to solve the optimization problem efficiently. Numerical results based on real-world datasets verify the effectiveness of the proposed DP-PASGD scheme.

We present a novel approach called Optimized Directed Roadmap Graph (ODRM). It is a method to build a directed roadmap graph that allows for collision avoidance in multi-robot navigation. This is a highly relevant problem, for example for industrial autonomous guided vehicles. The core idea of ODRM is, that a directed roadmap can encode inherent properties of the environment which are useful when agents have to avoid each other in that same environment. Like Probabilistic Roadmaps (PRMs), ODRM's first step is generating samples from C-space. In a second step, ODRM optimizes vertex positions and edge directions by Stochastic Gradient Descent (SGD). This leads to emergent properties like edges parallel to walls and patterns similar to two-lane streets or roundabouts. Agents can then navigate on this graph by searching their path independently and solving occurring agent-agent collisions at run-time. Using the graphs generated by ODRM compared to a non-optimized graph significantly fewer agent-agent collisions happen. We evaluate our roadmap with both, centralized and decentralized planners. Our experiments show that with ODRM even a simple centralized planner can solve problems with high numbers of agents that other multi-agent planners can not solve. Additionally, we use simulated robots with decentralized planners and online collision avoidance to show how agents are a lot faster on our roadmap than on standard grid maps.

In this paper, we propose a method of distributed stochastic gradient descent (SGD), with low communication load and computational complexity, and still fast convergence. To reduce the communication load, at each iteration of the algorithm, the worker nodes calculate and communicate some scalers, that are the directional derivatives of the sample functions in some \emph{pre-shared directions}. However, to maintain accuracy, after every specific number of iterations, they communicate the vectors of stochastic gradients. To reduce the computational complexity in each iteration, the worker nodes approximate the directional derivatives with zeroth-order stochastic gradient estimation, by performing just two function evaluations rather than computing a first-order gradient vector. The proposed method highly improves the convergence rate of the zeroth-order methods, guaranteeing order-wise faster convergence. Moreover, compared to the famous communication-efficient methods of model averaging (that perform local model updates and periodic communication of the gradients to synchronize the local models), we prove that for the general class of non-convex stochastic problems and with reasonable choice of parameters, the proposed method guarantees the same orders of communication load and convergence rate, while having order-wise less computational complexity. Experimental results on various learning problems in neural networks applications demonstrate the effectiveness of the proposed approach compared to various state-of-the-art distributed SGD methods.

This paper introduces a negative margin loss to metric learning based few-shot learning methods. The negative margin loss significantly outperforms regular softmax loss, and achieves state-of-the-art accuracy on three standard few-shot classification benchmarks with few bells and whistles. These results are contrary to the common practice in the metric learning field, that the margin is zero or positive. To understand why the negative margin loss performs well for the few-shot classification, we analyze the discriminability of learned features w.r.t different margins for training and novel classes, both empirically and theoretically. We find that although negative margin reduces the feature discriminability for training classes, it may also avoid falsely mapping samples of the same novel class to multiple peaks or clusters, and thus benefit the discrimination of novel classes.