Transformer-based architectures have been the subject of research aimed at understanding their overparameterization and the non-uniform importance of their layers. Applying these approaches to Automatic Speech Recognition, we demonstrate that the state-of-the-art Conformer models generally have multiple ambient layers. We study the stability of these layers across runs and model sizes, propose that group normalization may be used without disrupting their formation, and examine their correlation with model weight updates in each layer. Finally, we apply these findings to Federated Learning in order to improve the training procedure, by targeting Federated Dropout to layers by importance. This allows us to reduce the model size optimized by clients without quality degradation, and shows potential for future exploration.
相關內容
Human pose information is a critical component in many downstream image processing tasks, such as activity recognition and motion tracking. Likewise, a pose estimator for the illustrated character domain would provide a valuable prior for assistive content creation tasks, such as reference pose retrieval and automatic character animation. But while modern data-driven techniques have substantially improved pose estimation performance on natural images, little work has been done for illustrations. In our work, we bridge this domain gap by efficiently transfer-learning from both domain-specific and task-specific source models. Additionally, we upgrade and expand an existing illustrated pose estimation dataset, and introduce two new datasets for classification and segmentation subtasks. We then apply the resultant state-of-the-art character pose estimator to solve the novel task of pose-guided illustration retrieval. All data, models, and code will be made publicly available.
This work investigates the usage of batch normalization in neural architecture search (NAS). Specifically, Frankle et al. find that training BatchNorm only can achieve nontrivial performance. Furthermore, Chen et al. claim that training BatchNorm only can speed up the training of the one-shot NAS supernet over ten times. Critically, there is no effort to understand 1) why training BatchNorm only can find the perform-well architectures with the reduced supernet-training time, and 2) what is the difference between the train-BN-only supernet and the standard-train supernet. We begin by showing that the train-BN-only networks converge to the neural tangent kernel regime, obtain the same training dynamics as train all parameters theoretically. Our proof supports the claim to train BatchNorm only on supernet with less training time. Then, we empirically disclose that train-BN-only supernet provides an advantage on convolutions over other operators, cause unfair competition between architectures. This is due to only the convolution operator being attached with BatchNorm. Through experiments, we show that such unfairness makes the search algorithm prone to select models with convolutions. To solve this issue, we introduce fairness in the search space by placing a BatchNorm layer on every operator. However, we observe that the performance predictor in Chen et al. is inapplicable on the new search space. To this end, we propose a novel composite performance indicator to evaluate networks from three perspectives: expressivity, trainability, and uncertainty, derived from the theoretical property of BatchNorm. We demonstrate the effectiveness of our approach on multiple NAS-benchmarks (NAS-Bench101, NAS-Bench-201) and search spaces (DARTS search space and MobileNet search space).
Deep Learning has revolutionized the fields of computer vision, natural language understanding, speech recognition, information retrieval and more. However, with the progressive improvements in deep learning models, their number of parameters, latency, resources required to train, etc. have all have increased significantly. Consequently, it has become important to pay attention to these footprint metrics of a model as well, not just its quality. We present and motivate the problem of efficiency in deep learning, followed by a thorough survey of the five core areas of model efficiency (spanning modeling techniques, infrastructure, and hardware) and the seminal work there. We also present an experiment-based guide along with code, for practitioners to optimize their model training and deployment. We believe this is the first comprehensive survey in the efficient deep learning space that covers the landscape of model efficiency from modeling techniques to hardware support. Our hope is that this survey would provide the reader with the mental model and the necessary understanding of the field to apply generic efficiency techniques to immediately get significant improvements, and also equip them with ideas for further research and experimentation to achieve additional gains.
Deep graph neural networks (GNNs) have achieved excellent results on various tasks on increasingly large graph datasets with millions of nodes and edges. However, memory complexity has become a major obstacle when training deep GNNs for practical applications due to the immense number of nodes, edges, and intermediate activations. To improve the scalability of GNNs, prior works propose smart graph sampling or partitioning strategies to train GNNs with a smaller set of nodes or sub-graphs. In this work, we study reversible connections, group convolutions, weight tying, and equilibrium models to advance the memory and parameter efficiency of GNNs. We find that reversible connections in combination with deep network architectures enable the training of overparameterized GNNs that significantly outperform existing methods on multiple datasets. Our models RevGNN-Deep (1001 layers with 80 channels each) and RevGNN-Wide (448 layers with 224 channels each) were both trained on a single commodity GPU and achieve an ROC-AUC of $87.74 \pm 0.13$ and $88.14 \pm 0.15$ on the ogbn-proteins dataset. To the best of our knowledge, RevGNN-Deep is the deepest GNN in the literature by one order of magnitude. Please visit our project website //www.deepgcns.org/arch/gnn1000 for more information.
One of the key steps in Neural Architecture Search (NAS) is to estimate the performance of candidate architectures. Existing methods either directly use the validation performance or learn a predictor to estimate the performance. However, these methods can be either computationally expensive or very inaccurate, which may severely affect the search efficiency and performance. Moreover, as it is very difficult to annotate architectures with accurate performance on specific tasks, learning a promising performance predictor is often non-trivial due to the lack of labeled data. In this paper, we argue that it may not be necessary to estimate the absolute performance for NAS. On the contrary, we may need only to understand whether an architecture is better than a baseline one. However, how to exploit this comparison information as the reward and how to well use the limited labeled data remains two great challenges. In this paper, we propose a novel Contrastive Neural Architecture Search (CTNAS) method which performs architecture search by taking the comparison results between architectures as the reward. Specifically, we design and learn a Neural Architecture Comparator (NAC) to compute the probability of candidate architectures being better than a baseline one. Moreover, we present a baseline updating scheme to improve the baseline iteratively in a curriculum learning manner. More critically, we theoretically show that learning NAC is equivalent to optimizing the ranking over architectures. Extensive experiments in three search spaces demonstrate the superiority of our CTNAS over existing methods.
Federated learning (FL) is an emerging, privacy-preserving machine learning paradigm, drawing tremendous attention in both academia and industry. A unique characteristic of FL is heterogeneity, which resides in the various hardware specifications and dynamic states across the participating devices. Theoretically, heterogeneity can exert a huge influence on the FL training process, e.g., causing a device unavailable for training or unable to upload its model updates. Unfortunately, these impacts have never been systematically studied and quantified in existing FL literature. In this paper, we carry out the first empirical study to characterize the impacts of heterogeneity in FL. We collect large-scale data from 136k smartphones that can faithfully reflect heterogeneity in real-world settings. We also build a heterogeneity-aware FL platform that complies with the standard FL protocol but with heterogeneity in consideration. Based on the data and the platform, we conduct extensive experiments to compare the performance of state-of-the-art FL algorithms under heterogeneity-aware and heterogeneity-unaware settings. Results show that heterogeneity causes non-trivial performance degradation in FL, including up to 9.2% accuracy drop, 2.32x lengthened training time, and undermined fairness. Furthermore, we analyze potential impact factors and find that device failure and participant bias are two potential factors for performance degradation. Our study provides insightful implications for FL practitioners. On the one hand, our findings suggest that FL algorithm designers consider necessary heterogeneity during the evaluation. On the other hand, our findings urge system providers to design specific mechanisms to mitigate the impacts of heterogeneity.
The growing energy and performance costs of deep learning have driven the community to reduce the size of neural networks by selectively pruning components. Similarly to their biological counterparts, sparse networks generalize just as well, if not better than, the original dense networks. Sparsity can reduce the memory footprint of regular networks to fit mobile devices, as well as shorten training time for ever growing networks. In this paper, we survey prior work on sparsity in deep learning and provide an extensive tutorial of sparsification for both inference and training. We describe approaches to remove and add elements of neural networks, different training strategies to achieve model sparsity, and mechanisms to exploit sparsity in practice. Our work distills ideas from more than 300 research papers and provides guidance to practitioners who wish to utilize sparsity today, as well as to researchers whose goal is to push the frontier forward. We include the necessary background on mathematical methods in sparsification, describe phenomena such as early structure adaptation, the intricate relations between sparsity and the training process, and show techniques for achieving acceleration on real hardware. We also define a metric of pruned parameter efficiency that could serve as a baseline for comparison of different sparse networks. We close by speculating on how sparsity can improve future workloads and outline major open problems in the field.
The Transformer is widely used in natural language processing tasks. To train a Transformer however, one usually needs a carefully designed learning rate warm-up stage, which is shown to be crucial to the final performance but will slow down the optimization and bring more hyper-parameter tunings. In this paper, we first study theoretically why the learning rate warm-up stage is essential and show that the location of layer normalization matters. Specifically, we prove with mean field theory that at initialization, for the original-designed Post-LN Transformer, which places the layer normalization between the residual blocks, the expected gradients of the parameters near the output layer are large. Therefore, using a large learning rate on those gradients makes the training unstable. The warm-up stage is practically helpful for avoiding this problem. On the other hand, our theory also shows that if the layer normalization is put inside the residual blocks (recently proposed as Pre-LN Transformer), the gradients are well-behaved at initialization. This motivates us to remove the warm-up stage for the training of Pre-LN Transformers. We show in our experiments that Pre-LN Transformers without the warm-up stage can reach comparable results with baselines while requiring significantly less training time and hyper-parameter tuning on a wide range of applications.
Graph Neural Networks (GNNs) are based on repeated aggregations of information across nodes' neighbors in a graph. However, because common neighbors are shared between different nodes, this leads to repeated and inefficient computations. We propose Hierarchically Aggregated computation Graphs (HAGs), a new GNN graph representation that explicitly avoids redundancy by managing intermediate aggregation results hierarchically, eliminating repeated computations and unnecessary data transfers in GNN training and inference. We introduce an accurate cost function to quantitatively evaluate the runtime performance of different HAGs and use a novel HAG search algorithm to find optimized HAGs. Experiments show that the HAG representation significantly outperforms the standard GNN graph representation by increasing the end-to-end training throughput by up to 2.8x and reducing the aggregations and data transfers in GNN training by up to 6.3x and 5.6x, while maintaining the original model accuracy.
End-to-end approaches have drawn much attention recently for significantly simplifying the construction of an automatic speech recognition (ASR) system. RNN transducer (RNN-T) is one of the popular end-to-end methods. Previous studies have shown that RNN-T is difficult to train and a very complex training process is needed for a reasonable performance. In this paper, we explore RNN-T for a Chinese large vocabulary continuous speech recognition (LVCSR) task and aim to simplify the training process while maintaining performance. First, a new strategy of learning rate decay is proposed to accelerate the model convergence. Second, we find that adding convolutional layers at the beginning of the network and using ordered data can discard the pre-training process of the encoder without loss of performance. Besides, we design experiments to find a balance among the usage of GPU memory, training circle and model performance. Finally, we achieve 16.9% character error rate (CER) on our test set which is 2% absolute improvement from a strong BLSTM CE system with language model trained on the same text corpus.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191