This article describes an efficient training method for online streaming attention-based encoder-decoder (AED) automatic speech recognition (ASR) systems. AED models have achieved competitive performance in offline scenarios by jointly optimizing all components. They have recently been extended to an online streaming framework via models such as monotonic chunkwise attention (MoChA). However, the elaborate attention calculation process is not robust for long-form speech utterances. Moreover, the sequence-level training objective and time-restricted streaming encoder cause a nonnegligible delay in token emission during inference. To address these problems, we propose CTC synchronous training (CTC-ST), in which CTC alignments are leveraged as a reference for token boundaries to enable a MoChA model to learn optimal monotonic input-output alignments. We formulate a purely end-to-end training objective to synchronize the boundaries of MoChA to those of CTC. The CTC model shares an encoder with the MoChA model to enhance the encoder representation. Moreover, the proposed method provides alignment information learned in the CTC branch to the attention-based decoder. Therefore, CTC-ST can be regarded as self-distillation of alignment knowledge from CTC to MoChA. Experimental evaluations on a variety of benchmark datasets show that the proposed method significantly reduces recognition errors and emission latency simultaneously. The robustness to long-form and noisy speech is also demonstrated. We compare CTC-ST with several methods that distill alignment knowledge from a hybrid ASR system and show that the CTC-ST can achieve a comparable tradeoff of accuracy and latency without relying on external alignment information. The best MoChA system shows recognition accuracy comparable to that of RNN-transducer (RNN-T) while achieving lower emission latency.
In recent years, Convolutional Neural Networks (CNNs) have enabled ubiquitous image processing applications. As such, CNNs require fast runtime (forward propagation) to process high-resolution visual streams in real time. This is still a challenging task even with state-of-the-art graphics and tensor processing units. The bottleneck in computational efficiency primarily occurs in the convolutional layers. Performing operations in the Fourier domain is a promising way to accelerate forward propagation since it transforms convolutions into elementwise multiplications, which are considerably faster to compute for large kernels. Furthermore, such computation could be implemented using an optical 4f system with orders of magnitude faster operation. However, a major challenge in using this spectral approach, as well as in an optical implementation of CNNs, is the inclusion of a nonlinearity between each convolutional layer, without which CNN performance drops dramatically. Here, we propose a Spectral CNN Linear Counterpart (SCLC) network architecture and develop a Knowledge Distillation (KD) approach to circumvent the need for a nonlinearity and successfully train such networks. While the KD approach is known in machine learning as an effective process for network pruning, we adapt the approach to transfer the knowledge from a nonlinear network (teacher) to a linear counterpart (student). We show that the KD approach can achieve performance that easily surpasses the standard linear version of a CNN and could approach the performance of the nonlinear network. Our simulations show that the possibility of increasing the resolution of the input image allows our proposed 4f optical linear network to perform more efficiently than a nonlinear network with the same accuracy on two fundamental image processing tasks: (i) object classification and (ii) semantic segmentation.
Knowledge distillation is classically a procedure where a neural network is trained on the output of another network along with the original targets in order to transfer knowledge between the architectures. The special case of self-distillation, where the network architectures are identical, has been observed to improve generalization accuracy. In this paper, we consider an iterative variant of self-distillation in a kernel regression setting, in which successive steps incorporate both model outputs and the ground-truth targets. This allows us to provide the first theoretical results on the importance of using the weighted ground-truth targets in self-distillation. Our focus is on fitting nonlinear functions to training data with a weighted mean square error objective function suitable for distillation, subject to $\ell_2$ regularization of the model parameters. We show that any such function obtained with self-distillation can be calculated directly as a function of the initial fit, and that infinite distillation steps yields the same optimization problem as the original with amplified regularization. Furthermore, we provide a closed form solution for the optimal choice of weighting parameter at each step, and show how to efficiently estimate this weighting parameter for deep learning and significantly reduce the computational requirements compared to a grid search.
Knowledge distillation (KD) has witnessed its powerful capability in learning compact models in object detection. Previous KD methods for object detection mostly focus on imitating deep features within the imitation regions instead of mimicking classification logit due to its inefficiency in distilling localization information and trivial improvement. In this paper, by reformulating the knowledge distillation process on localization, we present a novel localization distillation (LD) method which can efficiently transfer the localization knowledge from the teacher to the student. Moreover, we also heuristically introduce the concept of valuable localization region that can aid to selectively distill the semantic and localization knowledge for a certain region. Combining these two new components, for the first time, we show that logit mimicking can outperform feature imitation and localization knowledge distillation is more important and efficient than semantic knowledge for distilling object detectors. Our distillation scheme is simple as well as effective and can be easily applied to different dense object detectors. Experiments show that our LD can boost the AP score of GFocal-ResNet-50 with a single-scale 1x training schedule from 40.1 to 42.1 on the COCO benchmark without any sacrifice on the inference speed. Our source code and trained models are publicly available at https://github.com/HikariTJU/LD
Knowledge distillation refers to a technique of transferring the knowledge from a large learned model or an ensemble of learned models to a small model. This method relies on access to the original training set, which might not always be available. A possible solution is a data-free adversarial distillation framework, which deploys a generative network to transfer the teacher model's knowledge to the student model. However, the data generation efficiency is low in the data-free adversarial distillation. We add an activation regularizer and a virtual interpolation method to improve the data generation efficiency. The activation regularizer enables the students to match the teacher's predictions close to activation boundaries and decision boundaries. The virtual interpolation method can generate virtual samples and labels in-between decision boundaries. Our experiments show that our approach surpasses state-of-the-art data-free distillation methods. The student model can achieve 95.42% accuracy on CIFAR-10 and 77.05% accuracy on CIFAR-100 without any original training data. Our model's accuracy is 13.8% higher than the state-of-the-art data-free method on CIFAR-100.
Multi-lingual script identification is a difficult task consisting of different language with complex backgrounds in scene text images. According to the current research scenario, deep neural networks are employed as teacher models to train a smaller student network by utilizing the teacher model's predictions. This process is known as dark knowledge transfer. It has been quite successful in many domains where the final result obtained is unachievable through directly training the student network with a simple architecture. In this paper, we explore dark knowledge transfer approach using long short-term memory(LSTM) and CNN based assistant model and various deep neural networks as the teacher model, with a simple CNN based student network, in this domain of multi-script identification from natural scene text images. We explore the performance of different teacher models and their ability to transfer knowledge to a student network. Although the small student network's limited size, our approach obtains satisfactory results on a well-known script identification dataset CVSI-2015.
Medical image segmentation models are typically supervised by expert annotations at the pixel-level, which can be expensive to acquire. In this work, we propose a method that combines the high quality of pixel-level expert annotations with the scale of coarse DNN-generated saliency maps for training multi-label semantic segmentation models. We demonstrate the application of our semi-supervised method, which we call CheXseg, on multi-label chest X-ray interpretation. We find that CheXseg improves upon the performance (mIoU) of fully-supervised methods that use only pixel-level expert annotations by 9.7% and weakly-supervised methods that use only DNN-generated saliency maps by 73.1%. Our best method is able to match radiologist agreement on three out of ten pathologies and reduces the overall performance gap by 57.2% as compared to weakly-supervised methods.
Neural dialogue models suffer from low-quality responses when interacted in practice, demonstrating difficulty in generalization beyond training data. Recently, knowledge distillation has been used to successfully regularize the student by transferring knowledge from the teacher. However, the teacher and the student are trained on the same dataset and tend to learn similar feature representations, whereas the most general knowledge should be found through differences. The finding of general knowledge is further hindered by the unidirectional distillation, as the student should obey the teacher and may discard some knowledge that is truly general but refuted by the teacher. To this end, we propose a novel training framework, where the learning of general knowledge is more in line with the idea of reaching consensus, i.e., finding common knowledge that is beneficial to different yet all datasets through diversified learning partners. Concretely, the training task is divided into a group of subtasks with the same number of students. Each student assigned to one subtask not only is optimized on the allocated subtask but also imitates multi-view feature representation aggregated from other students (i.e., student peers), which induces students to capture common knowledge among different subtasks and alleviates the over-fitting of students on the allocated subtasks. To further enhance generalization, we extend the unidirectional distillation to the bidirectional distillation that encourages the student and its student peers to co-evolve by exchanging complementary knowledge with each other. Empirical results and analysis demonstrate that our training framework effectively improves the model generalization without sacrificing training efficiency.
This paper is the first study to apply deep mutual learning (DML) to end-to-end ASR models. In DML, multiple models are trained simultaneously and collaboratively by mimicking each other throughout the training process, which helps to attain the global optimum and prevent models from making over-confident predictions. While previous studies applied DML to simple multi-class classification problems, there are no studies that have used it on more complex sequence-to-sequence mapping problems. For this reason, this paper presents a method to apply DML to state-of-the-art Transformer-based end-to-end ASR models. In particular, we propose to combine DML with recent representative training techniques. i.e., label smoothing, scheduled sampling, and SpecAugment, each of which are essential for powerful end-to-end ASR models. We expect that these training techniques work well with DML because DML has complementary characteristics. We experimented with two setups for Japanese ASR tasks: large-scale modeling and compact modeling. We demonstrate that DML improves the ASR performance of both modeling setups compared with conventional learning methods including knowledge distillation. We also show that combining DML with the existing training techniques effectively improves ASR performance.
We present a novel large-context end-to-end automatic speech recognition (E2E-ASR) model and its effective training method based on knowledge distillation. Common E2E-ASR models have mainly focused on utterance-level processing in which each utterance is independently transcribed. On the other hand, large-context E2E-ASR models, which take into account long-range sequential contexts beyond utterance boundaries, well handle a sequence of utterances such as discourses and conversations. However, the transformer architecture, which has recently achieved state-of-the-art ASR performance among utterance-level ASR systems, has not yet been introduced into the large-context ASR systems. We can expect that the transformer architecture can be leveraged for effectively capturing not only input speech contexts but also long-range sequential contexts beyond utterance boundaries. Therefore, this paper proposes a hierarchical transformer-based large-context E2E-ASR model that combines the transformer architecture with hierarchical encoder-decoder based large-context modeling. In addition, in order to enable the proposed model to use long-range sequential contexts, we also propose a large-context knowledge distillation that distills the knowledge from a pre-trained large-context language model in the training phase. We evaluate the effectiveness of the proposed model and proposed training method on Japanese discourse ASR tasks.
We propose a novel knowledge distillation approach to facilitate the transfer of dark knowledge from a teacher to a student. Contrary to most of the existing methods that rely on effective training of student models given pretrained teachers, we aim to learn the teacher models that are friendly to students and, consequently, more appropriate for knowledge transfer. In other words, at the time of optimizing a teacher model, the proposed algorithm learns the student branches jointly to obtain student-friendly representations. Since the main goal of our approach lies in training teacher models and the subsequent knowledge distillation procedure is straightforward, most of the existing knowledge distillation methods can adopt this technique to improve the performance of diverse student models in terms of accuracy and convergence speed. The proposed algorithm demonstrates outstanding accuracy in several well-known knowledge distillation techniques with various combinations of teacher and student models even in the case that their architectures are heterogeneous and there is no prior knowledge about student models at the time of training teacher networks.
Adversarial attack is aimed at fooling the target classifier with imperceptible perturbation. Adversarial examples, which are carefully crafted with a malicious purpose, can lead to erroneous predictions, resulting in catastrophic accidents. To mitigate the effects of adversarial attacks, we propose a novel purification model called CAP-GAN. CAP-GAN takes account of the idea of pixel-level and feature-level consistency to achieve reasonable purification under cycle-consistent learning. Specifically, we utilize the guided attention module and knowledge distillation to convey meaningful information to the purification model. Once a model is fully trained, inputs would be projected into the purification model and transformed into clean-like images. We vary the capacity of the adversary to argue the robustness against various types of attack strategies. On the CIFAR-10 dataset, CAP-GAN outperforms other pre-processing based defenses under both black-box and white-box settings.
In pickup and delivery services, transaction classification based on customer provided free text is a challenging problem. It involves the association of a wide variety of customer inputs to a fixed set of categories while adapting to the various customer writing styles. This categorization is important for the business: it helps understand the market needs and trends, and also assist in building a personalized experience for different segments of the customers. Hence, it is vital to capture these category information trends at scale, with high precision and recall. In this paper, we focus on a specific use-case where a single category drives each transaction. We propose a cost-effective transaction classification approach based on semi-supervision and knowledge distillation frameworks. The approach identifies the category of a transaction using free text input given by the customer. We use weak labelling and notice that the performance gains are similar to that of using human-annotated samples. On a large internal dataset and on 20Newsgroup dataset, we see that RoBERTa performs the best for the categorization tasks. Further, using an ALBERT model (it has 33x fewer parameters vis-a-vis parameters of RoBERTa), with RoBERTa as the Teacher, we see a performance similar to that of RoBERTa and better performance over unadapted ALBERT. This framework, with ALBERT as a student and RoBERTa as teacher, is further referred to as R-ALBERT in this paper. The model is in production and is used by business to understand changing trends and take appropriate decisions.
Intent classification is a task in spoken language understanding. An intent classification system is usually implemented as a pipeline process, with a speech recognition module followed by text processing that classifies the intents. There are also studies of end-to-end system that takes acoustic features as input and classifies the intents directly. Such systems don't take advantage of relevant linguistic information, and suffer from limited training data. In this work, we propose a novel intent classification framework that employs acoustic features extracted from a pretrained speech recognition system and linguistic features learned from a pretrained language model. We use knowledge distillation technique to map the acoustic embeddings towards linguistic embeddings. We perform fusion of both acoustic and linguistic embeddings through cross-attention approach to classify intents. With the proposed method, we achieve 90.86% and 99.07% accuracy on ATIS and Fluent speech corpus, respectively.
Existing methods for distillation do not efficiently utilize the training data. This work presents a novel approach to perform distillation using only a subset of the training data, making it more data-efficient. For this purpose, the training of the teacher model is modified to include self-regulation wherein a sample in the training set is used for updating model parameters in the backward pass either if it is misclassified or the model is not confident enough in its prediction. This modification restricts the participation of samples, unlike the conventional training method. The number of times a sample participates in the self-regulated training process is a measure of its significance towards the model's knowledge. The significance values are used to weigh the losses incurred on the corresponding samples in the distillation process. This method is named significance-based distillation. Two other methods are proposed for comparison where the student model learns by distillation and incorporating self-regulation as the teacher model, either utilizing the significance information computed during the teacher's training or not. These methods are named hybrid and regulated distillations, respectively. Experiments on benchmark datasets show that the proposed methods achieve similar performance as other state-of-the-art methods for knowledge distillation while utilizing a significantly less number of samples.
Negative sampling is a limiting factor w.r.t. the generalization of metric-learned neural networks. We show that uniform negative sampling provides little information about the class boundaries and thus propose three novel techniques for efficient negative sampling: drawing negative samples from (1) the top-$k$ most semantically similar classes, (2) the top-$k$ most semantically similar samples and (3) interpolating between contrastive latent representations to create pseudo negatives. Our experiments on CIFAR-10, CIFAR-100 and Tiny-ImageNet-200 show that our proposed \textit{Semantically Conditioned Negative Sampling} and Latent Mixup lead to consistent performance improvements. In the standard supervised learning setting, on average we increase test accuracy by 1.52\% percentage points on CIFAR-10 across various network architectures. In the knowledge distillation setting, (1) the performance of student networks increase by 4.56\% percentage points on Tiny-ImageNet-200 and 3.29\% on CIFAR-100 over student networks trained with no teacher and (2) 1.23\% and 1.72\% respectively over a \textit{hard-to-beat} baseline (Hinton et al., 2015).
Pre-trained language models (PLMs) like BERT have made great progress in NLP. News articles usually contain rich textual information, and PLMs have the potentials to enhance news text modeling for various intelligent news applications like news recommendation and retrieval. However, most existing PLMs are in huge size with hundreds of millions of parameters. Many online news applications need to serve millions of users with low latency tolerance, which poses huge challenges to incorporating PLMs in these scenarios. Knowledge distillation techniques can compress a large PLM into a much smaller one and meanwhile keeps good performance. However, existing language models are pre-trained and distilled on general corpus like Wikipedia, which has some gaps with the news domain and may be suboptimal for news intelligence. In this paper, we propose NewsBERT, which can distill PLMs for efficient and effective news intelligence. In our approach, we design a teacher-student joint learning and distillation framework to collaboratively learn both teacher and student models, where the student model can learn from the learning experience of the teacher model. In addition, we propose a momentum distillation method by incorporating the gradients of teacher model into the update of student model to better transfer useful knowledge learned by the teacher model. Extensive experiments on two real-world datasets with three tasks show that NewsBERT can effectively improve the model performance in various intelligent news applications with much smaller models.
Knowledge distillation extracts general knowledge from a pre-trained teacher network and provides guidance to a target student network. Most studies manually tie intermediate features of the teacher and student, and transfer knowledge through pre-defined links. However, manual selection often constructs ineffective links that limit the improvement from the distillation. There has been an attempt to address the problem, but it is still challenging to identify effective links under practical scenarios. In this paper, we introduce an effective and efficient feature distillation method utilizing all the feature levels of the teacher without manually selecting the links. Specifically, our method utilizes an attention-based meta-network that learns relative similarities between features, and applies identified similarities to control distillation intensities of all possible pairs. As a result, our method determines competent links more efficiently than the previous approach and provides better performance on model compression and transfer learning tasks. Further qualitative analyses and ablative studies describe how our method contributes to better distillation. The implementation code is available at github.com/clovaai/attention-feature-distillation.
Inference attacks against Machine Learning (ML) models allow adversaries to learn sensitive information about training data, model parameters, etc. While researchers have studied, in depth, several kinds of attacks, they have done so in isolation. As a result, we lack a comprehensive picture of the risks caused by the attacks, e.g., the different scenarios they can be applied to, the common factors that influence their performance, the relationship among them, or the effectiveness of possible defenses. In this paper, we fill this gap by presenting a first-of-its-kind holistic risk assessment of different inference attacks against machine learning models. We concentrate on four attacks -- namely, membership inference, model inversion, attribute inference, and model stealing -- and establish a threat model taxonomy. Our extensive experimental evaluation, run on five model architectures and four image datasets, shows that the complexity of the training dataset plays an important role with respect to the attack's performance, while the effectiveness of model stealing and membership inference attacks are negatively correlated. We also show that defenses like DP-SGD and Knowledge Distillation can only mitigate some of the inference attacks. Our analysis relies on a modular re-usable software, ML-Doctor, which enables ML model owners to assess the risks of deploying their models, and equally serves as a benchmark tool for researchers and practitioners.
One of the major limitations of deep learning models is that they face catastrophic forgetting in an incremental learning scenario. There have been several approaches proposed to tackle the problem of incremental learning. Most of these methods are based on knowledge distillation and do not adequately utilize the information provided by older task models, such as uncertainty estimation in predictions. The predictive uncertainty provides the distributional information can be applied to mitigate catastrophic forgetting in a deep learning framework. In the proposed work, we consider a Bayesian formulation to obtain the data and model uncertainties. We also incorporate self-attention framework to address the incremental learning problem. We define distillation losses in terms of aleatoric uncertainty and self-attention. In the proposed work, we investigate different ablation analyses on these losses. Furthermore, we are able to obtain better results in terms of accuracy on standard benchmarks.