Incremental Learning (IL) allows AI systems to adapt to streamed data. Most existing algorithms make two strong hypotheses which reduce the realism of the incremental scenario: (1) new data are assumed to be readily annotated when streamed and (2) tests are run with balanced datasets while most real-life datasets are actually imbalanced. These hypotheses are discarded and the resulting challenges are tackled with a combination of active and imbalanced learning. We introduce sample acquisition functions which tackle imbalance and are compatible with IL constraints. We also consider IL as an imbalanced learning problem instead of the established usage of knowledge distillation against catastrophic forgetting. Here, imbalance effects are reduced during inference through class prediction scaling. Evaluation is done with four visual datasets and compares existing and proposed sample acquisition functions. Results indicate that the proposed contributions have a positive effect and reduce the gap between active and standard IL performance.
Interpretability and effectiveness are two essential and indispensable requirements for adopting machine learning methods in reality. In this paper, we propose a knowledge distillation based decision trees extension, dubbed rectified decision trees (ReDT), to explore the possibility of fulfilling those requirements simultaneously. Specifically, we extend the splitting criteria and the ending condition of the standard decision trees, which allows training with soft labels while preserving the deterministic splitting paths. We then train the ReDT based on the soft label distilled from a well-trained teacher model through a novel jackknife-based method. Accordingly, ReDT preserves the excellent interpretable nature of the decision trees while having a relatively good performance. The effectiveness of adopting soft labels instead of hard ones is also analyzed empirically and theoretically. Surprisingly, experiments indicate that the introduction of soft labels also reduces the model size compared with the standard decision trees from the aspect of the total nodes and rules, which is an unexpected gift from the `dark knowledge' distilled from the teacher model.
Pretrained transformer models, such as BERT and T5, have shown to be highly effective at ad-hoc passage and document ranking. Due to inherent sequence length limits of these models, they need to be run over a document's passages, rather than processing the entire document sequence at once. Although several approaches for aggregating passage-level signals have been proposed, there has yet to be an extensive comparison of these techniques. In this work, we explore strategies for aggregating relevance signals from a document's passages into a final ranking score. We find that passage representation aggregation techniques can significantly improve over techniques proposed in prior work, such as taking the maximum passage score. We call this new approach PARADE. In particular, PARADE can significantly improve results on collections with broad information needs where relevance signals can be spread throughout the document (such as TREC Robust04 and GOV2). Meanwhile, less complex aggregation techniques may work better on collections with an information need that can often be pinpointed to a single passage (such as TREC DL and TREC Genomics). We also conduct efficiency analyses, and highlight several strategies for improving transformer-based aggregation.
Knowledge Distillation (KD) based methods adopt the one-way Knowledge Transfer (KT) scheme in which training a lower-capacity student network is guided by a pre-trained high-capacity teacher network. Recently, Deep Mutual Learning (DML) presented a two-way KT strategy, showing that the student network can be also helpful to improve the teacher network. In this paper, we propose Dense Cross-layer Mutual-distillation (DCM), an improved two-way KT method in which the teacher and student networks are trained collaboratively from scratch. To augment knowledge representation learning, well-designed auxiliary classifiers are added to certain hidden layers of both teacher and student networks. To boost KT performance, we introduce dense bidirectional KD operations between the layers appended with classifiers. After training, all auxiliary classifiers are discarded, and thus there are no extra parameters introduced to final models. We test our method on a variety of KT tasks, showing its superiorities over related methods. Code is available at https://github.com/sundw2014/DCM
In this paper, we propose an approach for filter-level pruning with hierarchical knowledge distillation based on the teacher, teaching-assistant, and student framework. Our method makes use of teaching assistants at intermediate pruning levels that share the same architecture and weights as the target student. We propose to prune each model independently using the gradient information from its corresponding teacher. By considering the relative sizes of each student-teacher pair, this formulation provides a natural trade-off between the capacity gap for knowledge distillation and the bias of the filter saliency updates. Our results show improvements in the attainable accuracy and model compression across the CIFAR10 and ImageNet classification tasks using the VGG16and ResNet50 architectures. We provide an extensive evaluation that demonstrates the benefits of using a varying number of teaching assistant models at different sizes.
The focus of this paper is dynamic gesture recognition in the context of the interaction between humans and machines. We propose a model consisting of two sub-networks, a transformer and an ordered-neuron long-short-term-memory (ON-LSTM) based recurrent neural network (RNN). Each sub-network is trained to perform the task of gesture recognition using only skeleton joints. Since each sub-network extracts different types of features due to the difference in architecture, the knowledge can be shared between the sub-networks. Through knowledge distillation, the features and predictions from each sub-network are fused together into a new fusion classifier. In addition, a cyclical learning rate can be used to generate a series of models that are combined in an ensemble, in order to yield a more generalizable prediction. The proposed ensemble of knowledge-sharing models exhibits an overall accuracy of 86.11% using only skeleton information, as tested using the Dynamic Hand Gesture-14/28 dataset
Deep neural networks (DNN) have recently been widely used in speaker recognition systems, achieving state-of-the-art performance on various benchmarks. The x-vector architecture is especially popular in this research community, due to its excellent performance and manageable computational complexity. In this paper, we present the lrx-vector system, which is the low-rank factorized version of the x-vector embedding network. The primary objective of this topology is to further reduce the memory requirement of the speaker recognition system. We discuss the deployment of knowledge distillation for training the lrx-vector system and compare against low-rank factorization with SVD. On the VOiCES 2019 far-field corpus we were able to reduce the weights by 28% compared to the full-rank x-vector system while keeping the recognition rate constant (1.83% EER).
Unsupervised crowd counting is a challenging yet not largely explored task. In this paper, we explore it in a transfer learning setting where we learn to detect and count persons in an unlabeled target set by transferring bi-knowledge learnt from regression- and detection-based models in a labeled source set. The dual source knowledge of the two models is heterogeneous and complementary as they capture different modalities of the crowd distribution. We formulate the mutual transformations between the outputs of regression- and detection-based models as two scene-agnostic transformers which enable knowledge distillation between the two models. Given the regression- and detection-based models and their mutual transformers learnt in the source, we introduce an iterative self-supervised learning scheme with regression-detection bi-knowledge transfer in the target. Extensive experiments on standard crowd counting benchmarks, ShanghaiTech, UCF\_CC\_50, and UCF\_QNRF demonstrate a substantial improvement of our method over other state-of-the-arts in the transfer learning setting.
In recent years, the fields of natural language processing (NLP) and information retrieval (IR) have made tremendous progress thanksto deep learning models like Recurrent Neural Networks (RNNs), Gated Recurrent Units (GRUs) and Long Short-Term Memory (LSTMs)networks, and Transformer [120] based models like Bidirectional Encoder Representations from Transformers (BERT) [24], GenerativePre-training Transformer (GPT-2) [94], Multi-task Deep Neural Network (MT-DNN) [73], Extra-Long Network (XLNet) [134], Text-to-text transfer transformer (T5) [95], T-NLG [98] and GShard [63]. But these models are humongous in size. On the other hand,real world applications demand small model size, low response times and low computational power wattage. In this survey, wediscuss six different types of methods (Pruning, Quantization, Knowledge Distillation, Parameter Sharing, Tensor Decomposition, andSub-quadratic Transformer based methods) for compression of such models to enable their deployment in real industry NLP projects.Given the critical need of building applications with efficient and small models, and the large amount of recently published work inthis area, we believe that this survey organizes the plethora of work done by the 'deep learning for NLP' community in the past fewyears and presents it as a coherent story.
Open set classification (OSC) tackles the problem of determining whether the data are in-class or out-of-class during inference, when only provided with a set of in-class examples at training time. Traditional OSC methods usually train discriminative or generative models with in-class data, then utilize the pre-trained models to classify test data directly. However, these methods always suffer from embedding confusion problem, i.e., partial out-of-class instances are mixed with in-class ones of similar semantics, making it difficult to classify. To solve this problem, we unify semi-supervised learning to develop a novel OSC algorithm, S2OSC, that incorporates out-of-class instances filtering and model re-training in a transductive manner. In detail, given a pool of newly coming test data, S2OSC firstly filters distinct out-of-class instances using the pre-trained model, and annotates super-class for them. Then, S2OSC trains a holistic classification model by combing in-class and out-of-class labeled data and remaining unlabeled test data in semi-supervised paradigm, which also integrates pre-trained model for knowledge distillation to further separate mixed instances. Despite its simplicity, the experimental results show that S2OSC achieves state-of-the-art performance across a variety of OSC tasks, including 85.4% of F1 on CIFAR-10 with only 300 pseudo-labels. We also demonstrate how S2OSC can be expanded to incremental OSC setting effectively with streaming data.
Attention-based sequence-to-sequence (seq2seq) models have achieved promising results in automatic speech recognition (ASR). However, as these models decode in a left-to-right way, they do not have access to context on the right. We leverage both left and right context by applying BERT as an external language model to seq2seq ASR through knowledge distillation. In our proposed method, BERT generates soft labels to guide the training of seq2seq ASR. Furthermore, we leverage context beyond the current utterance as input to BERT. Experimental evaluations show that our method significantly improves the ASR performance from the seq2seq baseline on the Corpus of Spontaneous Japanese (CSJ). Knowledge distillation from BERT outperforms that from a transformer LM that only looks at left context. We also show the effectiveness of leveraging context beyond the current utterance. Our method outperforms other LM application approaches such as n-best rescoring and shallow fusion, while it does not require extra inference cost.
Speech synthesis (text to speech, TTS) and recognition (automatic speech recognition, ASR) are important speech tasks, and require a large amount of text and speech pairs for model training. However, there are more than 6,000 languages in the world and most languages are lack of speech training data, which poses significant challenges when building TTS and ASR systems for extremely low-resource languages. In this paper, we develop LRSpeech, a TTS and ASR system under the extremely low-resource setting, which can support rare languages with low data cost. LRSpeech consists of three key techniques: 1) pre-training on rich-resource languages and fine-tuning on low-resource languages; 2) dual transformation between TTS and ASR to iteratively boost the accuracy of each other; 3) knowledge distillation to customize the TTS model on a high-quality target-speaker voice and improve the ASR model on multiple voices. We conduct experiments on an experimental language (English) and a truly low-resource language (Lithuanian) to verify the effectiveness of LRSpeech. Experimental results show that LRSpeech 1) achieves high quality for TTS in terms of both intelligibility (more than 98% intelligibility rate) and naturalness (above 3.5 mean opinion score (MOS)) of the synthesized speech, which satisfy the requirements for industrial deployment, 2) achieves promising recognition accuracy for ASR, and 3) last but not least, uses extremely low-resource training data. We also conduct comprehensive analyses on LRSpeech with different amounts of data resources, and provide valuable insights and guidances for industrial deployment. We are currently deploying LRSpeech into a commercialized cloud speech service to support TTS on more rare languages.
Semi-supervised learning (SSL) is an active area of research which aims to utilize unlabelled data in order to improve the accuracy of speech recognition systems. The current study proposes a methodology for integration of two key ideas: 1) SSL using connectionist temporal classification (CTC) objective and teacher-student based learning 2) Designing effective data-selection mechanisms for leveraging unlabelled data to boost performance of student models. Our aim is to establish the importance of good criteria in selecting samples from a large pool of unlabelled data based on attributes like confidence measure, speaker and content variability. The question we try to answer is: Is it possible to design a data selection mechanism which reduces dependence on a large set of randomly selected unlabelled samples without compromising on Word Error Rate (WER)? We perform empirical investigations of different data selection methods to answer this question and quantify the effect of different sampling strategies. On a semi-supervised ASR setting with 40000 hours of carefully selected unlabelled data, our CTC-SSL approach gives 17% relative WER improvement over a baseline CTC system trained with labelled data. It also achieves on-par performance with CTC-SSL system trained on order of magnitude larger unlabeled data based on random sampling.
Deep learning methods usually require a large amount of training data and lack interpretability. In this paper, we propose a novel knowledge distillation and model interpretation framework for medical image classification that jointly solves the above two issues. Specifically, to address the data-hungry issue, a small student model is learned with less data by distilling knowledge from a cumbersome pretrained teacher model. To interpret the teacher model and assist the learning of the student, an explainer module is introduced to highlight the regions of an input that are important for the predictions of the teacher model. Furthermore, the joint framework is trained by a principled way derived from the information-theoretic perspective. Our framework outperforms on the knowledge distillation and model interpretation tasks compared to state-of-the-art methods on a fundus dataset.
Knowledge distillation(KD) aims to improve the performance of a student network by mimicing the knowledge from a powerful teacher network. Existing methods focus on studying what knowledge should be transferred and treat all samples equally during training. This paper introduces the adaptive sample weighting to KD. We discover that previous effective hard mining methods are not appropriate for distillation. Furthermore, we propose Prime-Aware Adaptive Distillation (PAD) by the incorporation of uncertainty learning. PAD perceives the prime samples in distillation and then emphasizes their effect adaptively. PAD is fundamentally different from and would refine existing methods with the innovative view of unequal training. For this reason, PAD is versatile and has been applied in various tasks including classification, metric learning, and object detection. With ten teacher-student combinations on six datasets, PAD promotes the performance of existing distillation methods and outperforms recent state-of-the-art methods.
Convolutional Neural Networks (CNNs) are prone to overfit small training datasets. We present a novel two-phase pipeline that leverages self-supervised learning and knowledge distillation to improve the generalization ability of CNN models for image classification under the data-deficient setting. The first phase is to learn a teacher model which possesses rich and generalizable visual representations via self-supervised learning, and the second phase is to distill the representations into a student model in a self-distillation manner, and meanwhile fine-tune the student model for the image classification task. We also propose a novel margin loss for the self-supervised contrastive learning proxy task to better learn the representation under the data-deficient scenario. Together with other tricks, we achieve competitive performance in the VIPriors image classification challenge.
In this paper, we study the task of facial expression recognition under strong occlusion. We are particularly interested in cases where 50% of the face is occluded, e.g. when the subject wears a Virtual Reality (VR) headset. While previous studies show that pre-training convolutional neural networks (CNNs) on fully-visible (non-occluded) faces improves the accuracy, we propose to employ knowledge distillation to achieve further improvements. First of all, we employ the classic teacher-student training strategy, in which the teacher is a CNN trained on fully-visible faces and the student is a CNN trained on occluded faces. Second of all, we propose a new approach for knowledge distillation based on triplet loss. During training, the goal is to reduce the distance between an anchor embedding, produced by a student CNN that takes occluded faces as input, and a positive embedding (from the same class as the anchor), produced by a teacher CNN trained on fully-visible faces, so that it becomes smaller than the distance between the anchor and a negative embedding (from a different class than the anchor), produced by the student CNN. Third of all, we propose to combine the distilled embeddings obtained through the classic teacher-student strategy and our novel teacher-student strategy based on triplet loss into a single embedding vector. We conduct experiments on two benchmarks, FER+ and AffectNet, with two CNN architectures, VGG-f and VGG-face, showing that knowledge distillation can bring significant improvements over the state-of-the-art methods designed for occluded faces in the VR setting.
Knowledge distillation has become increasingly important in model compression. It boosts the performance of a miniaturized student network with the supervision of the output distribution and feature maps from a sophisticated teacher network. Some recent works introduce multi-teacher distillation to provide more supervision to the student network. However, the effectiveness of multi-teacher distillation methods are accompanied by costly computation resources. To tackle with both the efficiency and the effectiveness of knowledge distillation, we introduce the feature aggregation to imitate the multi-teacher distillation in the single-teacher distillation framework by extracting informative supervision from multiple teacher feature maps. Specifically, we introduce DFA, a two-stage Differentiable Feature Aggregation search method that motivated by DARTS in neural architecture search, to efficiently find the aggregations. In the first stage, DFA formulates the searching problem as a bi-level optimization and leverages a novel bridge loss, which consists of a student-to-teacher path and a teacher-to-student path, to find appropriate feature aggregations. The two paths act as two players against each other, trying to optimize the unified architecture parameters to the opposite directions while guaranteeing both expressivity and learnability of the feature aggregation simultaneously. In the second stage, DFA performs knowledge distillation with the derived feature aggregation. Experimental results show that DFA outperforms existing methods on CIFAR-100 and CINIC-10 datasets under various teacher-student settings, verifying the effectiveness and robustness of the design.
Scaling up the convolutional neural network (CNN) size (e.g., width, depth, etc.) is known to effectively improve model accuracy. However, the large model size impedes training on resource-constrained edge devices. For instance, federated learning (FL) may place undue burden on the compute capability of edge nodes, even though there is a strong practical need for FL due to its privacy and confidentiality properties. To address the resource-constrained reality of edge devices, we reformulate FL as a group knowledge transfer training algorithm, called FedGKT. FedGKT designs a variant of the alternating minimization approach to train small CNNs on edge nodes and periodically transfer their knowledge by knowledge distillation to a large server-side CNN. FedGKT consolidates several advantages into a single framework: reduced demand for edge computation, lower communication bandwidth for large CNNs, and asynchronous training, all while maintaining model accuracy comparable to FedAvg. We train CNNs designed based on ResNet-56 and ResNet-110 using three distinct datasets (CIFAR-10, CIFAR-100, and CINIC-10) and their non-I.I.D. variants. Our results show that FedGKT can obtain comparable or even slightly higher accuracy than FedAvg. More importantly, FedGKT makes edge training affordable. Compared to the edge training using FedAvg, FedGKT demands 9 to 17 times less computational power (FLOPs) on edge devices and requires 54 to 105 times fewer parameters in the edge CNN. Our source code is released at FedML (https://fedml.ai).
A crucial task in scene understanding is 3D object detection, which aims to detect and localize the 3D bounding boxes of objects belonging to specific classes. Existing 3D object detectors heavily rely on annotated 3D bounding boxes during training, while these annotations could be expensive to obtain and only accessible in limited scenarios. Weakly supervised learning is a promising approach to reducing the annotation requirement, but existing weakly supervised object detectors are mostly for 2D detection rather than 3D. In this work, we propose VS3D, a framework for weakly supervised 3D object detection from point clouds without using any ground truth 3D bounding box for training. First, we introduce an unsupervised 3D proposal module that generates object proposals by leveraging normalized point cloud densities. Second, we present a cross-modal knowledge distillation strategy, where a convolutional neural network learns to predict the final results from the 3D object proposals by querying a teacher network pretrained on image datasets. Comprehensive experiments on the challenging KITTI dataset demonstrate the superior performance of our VS3D in diverse evaluation settings. The source code and pretrained models are publicly available at https://github.com/Zengyi-Qin/Weakly-Supervised-3D-Object-Detection.