Stop Wasting my FLOPS: Improving the Efficiency of Deep Learning Models

Deep neural networks have completely revolutionized the field of machinelearning by achieving state-of-the-art results on various tasks ranging fromcomputer vision to protein folding. However, their application is hindered bytheir large computational and memory requirements. In this thesis, we proposemethods for improving the efficiency of deep neural networks.Firstly, we tackle the sample inefficiency of neural network training with animportance sampling algorithm suitable for deep neural networks. This algorithmallows us to focus computation on datapoints that are going to provide usefulgradients for training our models and ignore the ones that will have negligiblegradients. We show that our algorithm can improve the performance of variousneural networks when compared to uniform sampling under a fixed computationalbudget.Secondly, we design a model that is suitable for processing large input imageswith a fraction of the computational and memory requirements of traditionalapproaches. We achieve this by sampling from a data-dependent attentiondistribution in order to only process a portion of the input in highresolution. We demonstrate that our model can learn both the attention and thefeatures in an end-to-end fashion using only single image-wise labels forsupervision.Subsequently, we shift our attention to transformer architectures and introducea kernelized formulation for self-attention that reduces its quadraticcomplexity to linear with respect to the input sequence's length. Furthermore,we uncover the relationship between autoregressive transformers and recurrentneural networks and show that our formulation enables up to 3 orders ofmagnitude faster autoregressive inference.Finally, we develop clustered, attention a method that can approximate softmaxtransformers with reduced computation. This is achieved by grouping elements ofthe input using clustering. We showcase that our formulation provides a bettertrade-off between performance and computation in comparison to the originaltransformer architecture. In addition, we demonstrate that clustered attentioncan approximate pretrained transformer models without any fine-tuning and withminimal loss in performance.

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

Masked Training of Neural Networks with Partial Gradients

Martin Jaggi, Amirkeivan Mohtashami, Sebastian Urban Stich

State-of-the-art training algorithms for deep learning models are based on stochastic gradient descent (SGD). Recently, many variations have been explored: perturbing parameters for better accuracy (such as in Extra-gradient), limiting SGD updates to a subset of parameters for increased efficiency (such as meProp) or a combination of both (such as Dropout). However, the convergence of these methods is often not studied in theory. We propose a unified theoretical framework to study such SGD variants-encompassing the aforementioned algorithms and additionally a broad variety of methods used for communication efficient training or model compression. Our insights can be used as a guide to improve the efficiency of such methods and facilitate generalization to new applications. As an example, we tackle the task of jointly training networks, a version of which (limited to sub-networks) is used to create Slimmable Networks. By training a low-rank Transformer jointly with a standard one we obtain superior performance than when it is trained separately.

JMLR-JOURNAL MACHINE LEARNING RESEARCH2022

Word Sequence Modeling using Deep Learning

Joël Yvon Roland Legrand

For a long time, natural language processing (NLP) has relied on generative models with task specific and manually engineered features. Recently, there has been a resurgence of interest for neural networks in the machine learning community, obtaining state-of-the-art results in various fields such as computer vision, speech processing and natural language processing. The central idea behind these approaches is to learn features and models simultaneously, in an end-to-end manner, and making as few assumptions as possible. In NLP, word embeddings, mapping words in a dictionary on a continuous low-dimensional vector space, have proven to be very efficient for a large variety of tasks while requiring almost no a-priori linguistic assumptions. In this thesis, we investigate continuous representations of segments in a sentence for the purpose of solving NLP tasks that involve complex sentence-level relationships. Our sequence modelling approach is based on neural networks and takes advantage of word embeddings. A first approach models words in context in the form of continuous vector representations which are used to solve the task of interest. With the use of a compositional procedure, allowing arbitrarily-sized segments to be compressed onto continuous vectors, the model is able to consider long-range word dependencies as well. We first validate our approach on the task of bilingual word alignment, consisting in finding word correspondences between a sentence in two different languages. Source and target words in context are modeled using convolutional neural networks, obtaining representations that are later used to compute alignment scores. An aggregation operation enables unsupervised training for this task. We show that our model outperforms a standard generative model. The model above is extended to tackle phrase prediction tasks where phrases rather than single words are to be tagged. These tasks have been typically cast as classic word tagging problems using special tagging schemes to identify the segments boundaries. The proposed neural model focuses on learning fixed-size representations of arbitrarily-sized chunks of words that are used to solve the tagging task. A compositional operation is introduced in this work for the purpose of computing these representations. We demonstrate the viability of the proposed representations by evaluating the approach on the multiwork expression tagging task. The remainder of this thesis addresses the task of syntactic constituency parsing which, as opposed to the above tasks, aims at producing a structured output, in the form of a tree, of an input sentence. Syntactic parsing is cast as multiple phrase prediction problems that are solved recursively in a greedy manner. An extension using recursive compositional vector representations, allowing for lexical infor- mation to be propagated from early stages, is explored as well. This approach is evaluated on a standard corpus obtaining performance comparable to generative models with much shorter computation time. Finally, morphological tags are included as additional features, using a similar composition procedure, to improve the parsing performance for morphologically rich languages. State-of-the-art results were obtained for these task and languages.