Statistical Estimation and Inference via Local SGD in Federated Learning

Federated Learning (FL) makes a large amount of edge computing devices (e.g., mobile phones) jointly learn a global model without data sharing. In FL, data are generated in a decentralized manner with high heterogeneity. This paper studies how to perform statistical estimation and inference in the federated setting. We analyze the so-called Local SGD, a multi-round estimation procedure that uses intermittent communication to improve communication efficiency. We first establish a {it functional central limit theorem} that shows the averaged iterates of Local SGD weakly converge to a rescaled Brownian motion. We next provide two iterative inference methods: the {it plug-in} and the {it random scaling}. Random scaling constructs an asymptotically pivotal statistic for inference by using the information along the whole Local SGD path. Both the methods are communication efficient and applicable to online data. Our theoretical and empirical results show that Local SGD simultaneously achieves both statistical efficiency and communication efficiency.

Privacy-Preserving Distributed SVD via Federated Power

Singular value decomposition (SVD) is one of the most fundamental tools in machine learning and statistics.The modern machine learning community usually assumes that data come from and belong to small-scale device users. The low communication and computation power of such devices, and the possible privacy breaches of users sensitive data make the computation of SVD challenging. Federated learning (FL) is a paradigm enabling a large number of devices to jointly learn a model in a communication-efficient way without data sharing. In the FL framework, we develop a class of algorithms called FedPower for the computation of partial SVD in the modern setting. Based on the well-known power method, the local devices alternate between multiple local power iterations and one global aggregation to improve communication efficiency. In the aggregation, we propose to weight each local eigenvector matrix with Orthogonal Procrustes Transformation (OPT). Considering the practical stragglers effect, the aggregation can be fully participated or partially participated, where for the latter we propose two sampling and aggregation schemes. Further, to ensure strong privacy protection, we add Gaussian noise whenever the communication happens by adopting the notion of differential privacy (DP). We theoretically show the convergence bound for FedPower. The resulting bound is interpretable with each part corresponding to the effect of Gaussian noise, parallelization, and random sampling of devices, respectively. We also conduct experiments to demonstrate the merits of FedPower. In particular, the local iterations not only improve communication efficiency but also reduce the chance of privacy breaches.

Provable Benefits of Overparameterization in Model Compression: From Double Descent to Pruning Neural Networks

Deep networks are typically trained with many more parameters than the size of the training dataset. Recent empirical evidence indicates that the practice of overparameterization not only benefits training large models, but also assists - perhaps counterintuitively - building lightweight models. Specifically, it suggests that overparameterization benefits model pruning / sparsification. This paper sheds light on these empirical findings by theoretically characterizing the high-dimensional asymptotics of model pruning in the overparameterized regime. The theory presented addresses the following core question: should one train a small model from the beginning, or first train a large model and then prune?. We analytically identify regimes in which, even if the location of the most informative features is known, we are better off fitting a large model and then pruning rather than simply training with the known informative features. This leads to a new double descent in the training of sparse models: growing the original model, while preserving the target sparsity, improves the test accuracy as one moves beyond the overparameterization threshold. Our analysis further reveals the benefit of retraining by relating it to feature correlations. We find that the above phenomena are already present in linear and random-features models. Our technical approach advances the toolset of high-dimensional analysis and precisely characterizes the asymptotic distribution of over-parameterized least-squares. The intuition gained by analytically studying simpler models is numerically verified on neural networks.

Kernel Interpolation of High Dimensional Scattered Data

Data sites selected from modeling high-dimensional problems often appear scattered in non-paternalistic ways. Except for sporadic-clustering at some spots, they become relatively far apart as the dimension of the ambient space grows. These features defy any theoretical treatment that requires local or global quasi-uniformity of distribution of data sites. Incorporating a recently-developed application of integral operator theory in machine learning, we propose and study in the current article a new framework to analyze kernel interpolation of high dimensional data, which features bounding stochastic approximation error by a hybrid (discrete and continuous) $K$-functional tied to the spectrum of the underlying kernel matrix. Both theoretical analysis and numerical simulations show that spectra of kernel matrices are reliable and stable barometers for gauging the performance of kernel-interpolation methods for high dimensional data.