Pruning aims to reduce the number of parameters while maintaining performance close to the original network. This work proposes a novel \emph{self-distillation} based pruning strategy, whereby the representational similarity between the pruned and unpruned versions of the same network is maximized. Unlike previous approaches that treat distillation and pruning separately, we use distillation to inform the pruning criteria, without requiring a separate student network as in knowledge distillation. We show that the proposed {\em cross-correlation objective for self-distilled pruning} implicitly encourages sparse solutions, naturally complementing magnitude-based pruning criteria. Experiments on the GLUE and XGLUE benchmarks show that self-distilled pruning increases mono- and cross-lingual language model performance. Self-distilled pruned models also outperform smaller Transformers with an equal number of parameters and are competitive against (6 times) larger distilled networks. We also observe that self-distillation (1) maximizes class separability, (2) increases the signal-to-noise ratio, and (3) converges faster after pruning steps, providing further insights into why self-distilled pruning improves generalization.
相關內容
Deep neural networks (DNNs) have recently achieved a great success in computer vision and several related fields. Despite such progress, current neural architectures still suffer from catastrophic interference (a.k.a. forgetting) which obstructs DNNs to learn continually. While several state-of-the-art methods have been proposed to mitigate forgetting, these existing solutions are either highly rigid (as regularization) or time/memory demanding (as replay). An intermediate class of methods, based on dynamic networks, has been proposed in the literature and provides a reasonable balance between task memorization and computational footprint. In this paper, we devise a dynamic network architecture for continual learning based on a novel forgetting-free neural block (FFNB). Training FFNB features on new tasks is achieved using a novel procedure that constrains the underlying parameters in the null-space of the previous tasks, while training classifier parameters equates to Fisher discriminant analysis. The latter provides an effective incremental process which is also optimal from a Bayesian perspective. The trained features and classifiers are further enhanced using an incremental "end-to-end" fine-tuning. Extensive experiments, conducted on different challenging classification problems, show the high effectiveness of the proposed method.
Most of the existing neural video compression methods adopt the predictive coding framework, which first generates the predicted frame and then encodes its residue with the current frame. However, as for compression ratio, predictive coding is only a sub-optimal solution as it uses simple subtraction operation to remove the redundancy across frames. In this paper, we propose a deep contextual video compression framework to enable a paradigm shift from predictive coding to conditional coding. In particular, we try to answer the following questions: how to define, use, and learn condition under a deep video compression framework. To tap the potential of conditional coding, we propose using feature domain context as condition. This enables us to leverage the high dimension context to carry rich information to both the encoder and the decoder, which helps reconstruct the high-frequency contents for higher video quality. Our framework is also extensible, in which the condition can be flexibly designed. Experiments show that our method can significantly outperform the previous state-of-the-art (SOTA) deep video compression methods. When compared with x265 using veryslow preset, we can achieve 26.0% bitrate saving for 1080P standard test videos.
In real-world applications, data often come in a growing manner, where the data volume and the number of classes may increase dynamically. This will bring a critical challenge for learning: given the increasing data volume or the number of classes, one has to instantaneously adjust the neural model capacity to obtain promising performance. Existing methods either ignore the growing nature of data or seek to independently search an optimal architecture for a given dataset, and thus are incapable of promptly adjusting the architectures for the changed data. To address this, we present a neural architecture adaptation method, namely Adaptation eXpert (AdaXpert), to efficiently adjust previous architectures on the growing data. Specifically, we introduce an architecture adjuster to generate a suitable architecture for each data snapshot, based on the previous architecture and the different extent between current and previous data distributions. Furthermore, we propose an adaptation condition to determine the necessity of adjustment, thereby avoiding unnecessary and time-consuming adjustments. Extensive experiments on two growth scenarios (increasing data volume and number of classes) demonstrate the effectiveness of the proposed method.
While pre-trained language models (e.g., BERT) have achieved impressive results on different natural language processing tasks, they have large numbers of parameters and suffer from big computational and memory costs, which make them difficult for real-world deployment. Therefore, model compression is necessary to reduce the computation and memory cost of pre-trained models. In this work, we aim to compress BERT and address the following two challenging practical issues: (1) The compression algorithm should be able to output multiple compressed models with different sizes and latencies, in order to support devices with different memory and latency limitations; (2) The algorithm should be downstream task agnostic, so that the compressed models are generally applicable for different downstream tasks. We leverage techniques in neural architecture search (NAS) and propose NAS-BERT, an efficient method for BERT compression. NAS-BERT trains a big supernet on a search space containing a variety of architectures and outputs multiple compressed models with adaptive sizes and latency. Furthermore, the training of NAS-BERT is conducted on standard self-supervised pre-training tasks (e.g., masked language model) and does not depend on specific downstream tasks. Thus, the compressed models can be used across various downstream tasks. The technical challenge of NAS-BERT is that training a big supernet on the pre-training task is extremely costly. We employ several techniques including block-wise search, search space pruning, and performance approximation to improve search efficiency and accuracy. Extensive experiments on GLUE and SQuAD benchmark datasets demonstrate that NAS-BERT can find lightweight models with better accuracy than previous approaches, and can be directly applied to different downstream tasks with adaptive model sizes for different requirements of memory or latency.
Understanding the inner workings of deep neural networks (DNNs) is essential to provide trustworthy artificial intelligence techniques for practical applications. Existing studies typically involve linking semantic concepts to units or layers of DNNs, but fail to explain the inference process. In this paper, we introduce neural architecture disentanglement (NAD) to fill the gap. Specifically, NAD learns to disentangle a pre-trained DNN into sub-architectures according to independent tasks, forming information flows that describe the inference processes. We investigate whether, where, and how the disentanglement occurs through experiments conducted with handcrafted and automatically-searched network architectures, on both object-based and scene-based datasets. Based on the experimental results, we present three new findings that provide fresh insights into the inner logic of DNNs. First, DNNs can be divided into sub-architectures for independent tasks. Second, deeper layers do not always correspond to higher semantics. Third, the connection type in a DNN affects how the information flows across layers, leading to different disentanglement behaviors. With NAD, we further explain why DNNs sometimes give wrong predictions. Experimental results show that misclassified images have a high probability of being assigned to task sub-architectures similar to the correct ones. Code will be available at: //github.com/hujiecpp/NAD.
Normalization is known to help the optimization of deep neural networks. Curiously, different architectures require specialized normalization methods. In this paper, we study what normalization is effective for Graph Neural Networks (GNNs). First, we adapt and evaluate the existing methods from other domains to GNNs. Faster convergence is achieved with InstanceNorm compared to BatchNorm and LayerNorm. We provide an explanation by showing that InstanceNorm serves as a preconditioner for GNNs, but such preconditioning effect is weaker with BatchNorm due to the heavy batch noise in graph datasets. Second, we show that the shift operation in InstanceNorm results in an expressiveness degradation of GNNs for highly regular graphs. We address this issue by proposing GraphNorm with a learnable shift. Empirically, GNNs with GraphNorm converge faster compared to GNNs using other normalization. GraphNorm also improves the generalization of GNNs, achieving better performance on graph classification benchmarks.
Since hardware resources are limited, the objective of training deep learning models is typically to maximize accuracy subject to the time and memory constraints of training and inference. We study the impact of model size in this setting, focusing on Transformer models for NLP tasks that are limited by compute: self-supervised pretraining and high-resource machine translation. We first show that even though smaller Transformer models execute faster per iteration, wider and deeper models converge in significantly fewer steps. Moreover, this acceleration in convergence typically outpaces the additional computational overhead of using larger models. Therefore, the most compute-efficient training strategy is to counterintuitively train extremely large models but stop after a small number of iterations. This leads to an apparent trade-off between the training efficiency of large Transformer models and the inference efficiency of small Transformer models. However, we show that large models are more robust to compression techniques such as quantization and pruning than small models. Consequently, one can get the best of both worlds: heavily compressed, large models achieve higher accuracy than lightly compressed, small models.
Deep convolutional neural networks (CNNs) have recently achieved great success in many visual recognition tasks. However, existing deep neural network models are computationally expensive and memory intensive, hindering their deployment in devices with low memory resources or in applications with strict latency requirements. Therefore, a natural thought is to perform model compression and acceleration in deep networks without significantly decreasing the model performance. During the past few years, tremendous progress has been made in this area. In this paper, we survey the recent advanced techniques for compacting and accelerating CNNs model developed. These techniques are roughly categorized into four schemes: parameter pruning and sharing, low-rank factorization, transferred/compact convolutional filters, and knowledge distillation. Methods of parameter pruning and sharing will be described at the beginning, after that the other techniques will be introduced. For each scheme, we provide insightful analysis regarding the performance, related applications, advantages, and drawbacks etc. Then we will go through a few very recent additional successful methods, for example, dynamic capacity networks and stochastic depths networks. After that, we survey the evaluation matrix, the main datasets used for evaluating the model performance and recent benchmarking efforts. Finally, we conclude this paper, discuss remaining challenges and possible directions on this topic.
With the widespread applications of deep convolutional neural networks (DCNNs), it becomes increasingly important for DCNNs not only to make accurate predictions but also to explain how they make their decisions. In this work, we propose a CHannel-wise disentangled InterPretation (CHIP) model to give the visual interpretation to the predictions of DCNNs. The proposed model distills the class-discriminative importance of channels in networks by utilizing the sparse regularization. Here, we first introduce the network perturbation technique to learn the model. The proposed model is capable to not only distill the global perspective knowledge from networks but also present the class-discriminative visual interpretation for specific predictions of networks. It is noteworthy that the proposed model is able to interpret different layers of networks without re-training. By combining the distilled interpretation knowledge in different layers, we further propose the Refined CHIP visual interpretation that is both high-resolution and class-discriminative. Experimental results on the standard dataset demonstrate that the proposed model provides promising visual interpretation for the predictions of networks in image classification task compared with existing visual interpretation methods. Besides, the proposed method outperforms related approaches in the application of ILSVRC 2015 weakly-supervised localization task.
Over the past few years, various tasks involving videos such as classification, description, summarization and question answering have received a lot of attention. Current models for these tasks compute an encoding of the video by treating it as a sequence of images and going over every image in the sequence. However, for longer videos this is very time consuming. In this paper, we focus on the task of video classification and aim to reduce the computational time by using the idea of distillation. Specifically, we first train a teacher network which looks at all the frames in a video and computes a representation for the video. We then train a student network whose objective is to process only a small fraction of the frames in the video and still produce a representation which is very close to the representation computed by the teacher network. This smaller student network involving fewer computations can then be employed at inference time for video classification. We experiment with the YouTube-8M dataset and show that the proposed student network can reduce the inference time by upto 30% with a very small drop in the performance
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