Quasi-Global Momentum: Accelerating Decentralized Deep Learning on Heterogeneous Data

Decentralized training of deep learning models is a key element for enabling data privacy and on-device learning over networks. In realistic learning scenarios, the presence of heterogeneity across different clients' local datasets poses an optimization challenge and may severely deteriorate the generalization performance. In this paper, we investigate and identify the limitation of several decentralized optimization algorithms for different degrees of data heterogeneity. We propose a novel momentum-based method to mitigate this decentralized training difficulty. We show in extensive empirical experiments on various CV/NLP datasets (CIFAR-10, ImageNet, and AG News) and several network topologies (Ring and Social Network) that our method is much more robust to the heterogeneity of clients' data than other existing methods, by a significant improvement in test performance (1% - 20%). Our code is publicly available(1).

Optimization methods for collaborative learning

Sai Praneeth Reddy Karimireddy

A traditional machine learning pipeline involves collecting massive amounts of data centrally on a server and training models to fit the data. However, increasing concerns about the privacy and security of user's data, combined with the sheer growth in the data sizes has incentivized looking beyond such traditional centralized approaches. Collaborative learning (which encompasses distributed, federated, and decentralized learning) proposes instead for a network of data holders to collaborate together to train models without transmitting any data. This new paradigm minimizes data exposure, but inherently faces some fundamental challenges. In this thesis, we bring to bear the framework of stochastic optimization to formalize and develop new algorithms for these challenges. This serves not only to develop novel solutions, but also to test the utility of the optimization lens in modern deep learning.We study three fundamental problems. Firstly, collaborative training replaces a one-time transmission of raw data with repeated rounds of communicating partially trained models. However, this quickly runs against bandwidth constraints when dealing with large models. We propose to solve this bandwidth constraint using compressed communication. Next, collaborative training leverages the computation power of the data holders directly. However, this is not as reliable as using a data center with only a subset of them available at any given time. Thus, we require new algorithms which can efficiently utilize unreliable local computation of the data holders. Finally, collaborative training allows any data holder to participate in the training process, without being able to inspect their data or local computation. This may potentially open the system to malicious or faulty agents who seek to derail the training. We develop algorithms with Byzantine robustness which are guaranteed to be resilient to such attackers.

EPFL2021

From Classical to Unsupervised-Deep-Learning Methods for Solving Inverse Problems in Imaging.

Harshit Gupta

In this thesis, we propose new algorithms to solve inverse problems in the context of biomedical images. Due to ill-posedness, solving these problems require some prior knowledge of the statistics of the underlying images. The traditional algorithms, in the field, assume prior knowledge related to smoothness or sparsity of these images. Recently, they have been outperformed by the second generation algorithms which harness the power of neural networks to learn required statistics from training data. Even more recently, last generation deep-learning-based methods have emerged which require neither training nor training data. This thesis devises algorithms which progress through these generations. It extends these generations to novel formulations and applications while bringing more robustness. In parallel, it also progresses in terms of complexity, from proposing algorithms for problems with 1D data and an exact known forward model to the ones with 4D data and an unknown parametric forward model. We introduce five main contributions. The last three of them propose deep-learning-based latest-generation algorithms that require no prior training. 1) We develop algorithms to solve the continuous-domain formulation of inverse problems with both classical Tikhonov and total-variation regularizations. We formalize the problems, characterize the solution set, and devise numerical approaches to find the solutions. 2) We propose an algorithm that improves upon end-to-end neural-network-based second generation algorithms. In our method, a neural network is first trained as a projector on a training set, and is then plugged in as a projector inside the projected gradient descent (PGD). Since the problem is nonconvex, we relax the PGD to ensure convergence to a local minimum under some constraints. This method outperforms all the previous generation algorithms for Computed Tomography (CT). 3) We develop a novel time-dependent deep-image-prior algorithm for modalities that involve a temporal sequence of images. We parameterize them as the output of an untrained neural network fed with a sequence of latent variables. To impose temporal directionality, the latent variables are assumed to lie on a 1D manifold. The network is then tuned to minimize the data fidelity. We obtain state-of-the-art results in dynamic magnetic resonance imaging (MRI) and even recover intra-frame images. 4) We propose a novel reconstruction paradigm for cryo-electron-microscopy (CryoEM) called CryoGAN. Motivated by generative adversarial networks (GANs), we reconstruct a biomolecule's 3D structure such that its CryoEM measurements resemble the acquired data in a distributional sense. The algorithm is pose-or-likelihood-estimation-free, needs no ab initio, and is proven to have a theoretical guarantee of recovery of the true structure. 5) We extend CryoGAN to reconstruct continuously varying conformations of a structure from heterogeneous data. We parameterize the conformations as the output of a neural network fed with latent variables on a low-dimensional manifold. The method is shown to recover continuous protein conformations and their energy landscape.

EPFL2020

Bridging the gap between model-driven and data-driven methods in the era of Big Data

Gael Lederrey

Data-driven and model-driven methodologies can be regarded as competitive fields since they tackle similar problems such as prediction. However, these two fields can learn from each other to improve themselves. Indeed, data-driven methodologies have been developed to use advanced methodologies based on Big Data technologies. On the other hand, model-driven methodologies concentrate on developing mathematical models based on theory and expert knowledge to allow for interpretability and control. Through three main contributions, this thesis aims to bridge the gap between these two fields by using their strengths and applying them to its counterpart.Discrete Choice Models (DCMs) have shown tremendous success in many fields, such as transportation. However, they have not evolved to tackle the growing amount of available data. On the other hand, Machine Learning (ML) researchers have developed optimization algorithms to efficiently estimate complex models on large datasets. Similarly, faster estimation of DCMs on larger datasets would improve the efficiency of modelers as well as enable new research axes. Thus, we take inspiration from the large body of existing research in efficient parameter estimation with extensive data and large numbers of parameters in deep learning and apply it to DCMs. The first chapter of this thesis introduces the HAMABS algorithm, which combines three fundamental principles to enable faster parameter estimation of DCMs (20x speedup compared to standard estimation) without compromising the precision of the parameter estimates.Collecting large amounts of data can be cumbersome and costly, even in the era of Big Data. For example, ML researchers in Computer Vision have been developing generative deep learning models to augment datasets. DCM researchers face similar issues with tabular data, e.g. travel surveys. In addition, if the collection process is not performed correctly, these datasets can contain bias, lack consistency, or be unrepresentative of the actual population. The second chapter of this thesis introduces the DATGAN, a Generative Adversarial Network (GAN) integrating expert knowledge to control the generation process. This new architecture allows modelers to generate controlled and representative synthetic data, outperforming similar state-of-the-art generative models. Finally, researchers are increasingly developing fully disaggregate agent-based simulation models, which use detailed synthetic populations to generate aggregate passenger flows. However, detailed disaggregate socioeconomic data is usually expensive to collect and heavily restricted in terms of access and usage. As such, synthetic populations are typically either drawn randomly from aggregate level control totals, limiting their quality, or tightly controlled, limiting their application and usefulness. To combat this, the third chapter extends the DATGAN methodology to generate highly detailed and consistent synthetic populations from small sample data. First, ciDATGAN learns to generate the variables in a low-sample highly detailed dataset, e.g. household travel survey. It then completes a high-sample dataset with few variables, e.g. microdata census, by generating the previously learned variables. The results show that this methodology can correct for bias and may enable the transfer of synthetic populations to new areas/contexts.

EPFL2022