Outlier detection is a key field of machine learning for identifying abnormal data objects. Due to the high expense of acquiring ground truth, unsupervised models are often chosen in practice. To compensate for the unstable nature of unsupervised algorithms, practitioners from high-stakes fields like finance, health, and security, prefer to build a large number of models for further combination and analysis. However, this poses scalability challenges in high-dimensional large datasets. In this study, we propose a three-module acceleration framework called SUOD to expedite the training and prediction with a large number of unsupervised detection models. SUOD's Random Projection module can generate lower subspaces for high-dimensional datasets while reserving their distance relationship. Balanced Parallel Scheduling module can forecast the training and prediction cost of models with high confidence---so the task scheduler could assign nearly equal amount of taskload among workers for efficient parallelization. SUOD also comes with a Pseudo-supervised Approximation module, which can approximate fitted unsupervised models by lower time complexity supervised regressors for fast prediction on unseen data. It may be considered as an unsupervised model knowledge distillation process. Notably, all three modules are independent with great flexibility to "mix and match"; a combination of modules can be chosen based on use cases. Extensive experiments on more than 30 benchmark datasets have shown the efficacy of SUOD, and a comprehensive future development plan is also presented.
Knowledge Distillation (KD) is a model-agnostic technique to improve model quality while having a fixed capacity budget. It is a commonly used technique for model compression, where a larger capacity teacher model with better quality is used to train a more compact student model with better inference efficiency. Through distillation, one hopes to benefit from student's compactness, without sacrificing too much on model quality. Despite the large success of knowledge distillation, better understanding of how it benefits student model's training dynamics remains under-explored. In this paper, we categorize teacher's knowledge into three hierarchical levels and study its effects on knowledge distillation: (1) knowledge of the `universe', where KD brings a regularization effect through label smoothing; (2) domain knowledge, where teacher injects class relationships prior to student's logit layer geometry; and (3) instance specific knowledge, where teacher rescales student model's per-instance gradients based on its measurement on the event difficulty. Using systematic analyses and extensive empirical studies on both synthetic and real-world datasets, we confirm that the aforementioned three factors play a major role in knowledge distillation. Furthermore, based on our findings, we diagnose some of the failure cases of applying KD from recent studies.
Convolutional neural networks (CNNs) are extensively beneficial for medical image processing. Medical images are plentiful, but there is a lack of annotated data. Transfer learning is used to solve the problem of lack of labeled data and grants CNNs better training capability. Transfer learning can be used in many different medical applications; however, the model under transfer should have the same size as the original network. Knowledge distillation is recently proposed to transfer the knowledge of a model to another one and can be useful to cover the shortcomings of transfer learning. But some parts of the knowledge may not be distilled by knowledge distillation. In this paper, a novel knowledge distillation using transfer learning is proposed to transfer the whole knowledge of a model to another one. The proposed method can be beneficial and practical for medical image analysis in which a small number of labeled data are available. The proposed process is tested for diabetic retinopathy classification. Simulation results demonstrate that using the proposed method, knowledge of an extensive network can be transferred to a smaller model.
The segmentation of brain tumors in multimodal MRIs is one of the most challenging tasks in medical image analysis. The recent state of the art algorithms solving this task is based on machine learning approaches and deep learning in particular. The amount of data used for training such models and its variability is a keystone for building an algorithm with high representation power. In this paper, we study the relationship between the performance of the model and the amount of data employed during the training process. On the example of brain tumor segmentation challenge, we compare the model trained with labeled data provided by challenge organizers, and the same model trained in omni-supervised manner using additional unlabeled data annotated with the ensemble of heterogeneous models. As a result, a single model trained with additional data achieves performance close to the ensemble of multiple models and outperforms individual methods.
In this paper, we propose a novel model compression approach to effectively compress BERT by progressive module replacing. Our approach first divides the original BERT into several modules and builds their compact substitutes. Then, we randomly replace the original modules with their substitutes to train the compact modules to mimic the behavior of the original modules. We progressively increase the probability of replacement through the training. In this way, our approach brings a deeper level of interaction between the original and compact models. Compared to the previous knowledge distillation approaches for BERT compression, our approach does not introduce any additional loss function. Our approach outperforms existing knowledge distillation approaches on GLUE benchmark, showing a new perspective of model compression.
Feature maps contain rich information about image intensity and spatial correlation. However, previous online knowledge distillation methods only utilize the class probabilities. Thus in this paper, we propose an online knowledge distillation method that transfers not only the knowledge of the class probabilities but also that of the feature map using the adversarial training framework. We train multiple networks simultaneously by employing discriminators to distinguish the feature map distributions of different networks. Each network has its corresponding discriminator which discriminates the feature map from its own as fake while classifying that of the other network as real. By training a network to fool the corresponding discriminator, it can learn the other network's feature map distribution. We show that our method performs better than the conventional direct alignment method such as L1 and is more suitable for online distillation. Also, we propose a novel cyclic learning scheme for training more than two networks together. We have applied our method to various network architectures on the classification task and discovered a significant improvement of performance especially in the case of training a pair of a small network and a large one.
Over the past year, the emergence of transfer learning with large-scale language models (LM) has led to dramatic performance improvements across a broad range of natural language understanding tasks. However, the size and memory footprint of these large LMs makes them difficult to deploy in many scenarios (e.g. on mobile phones). Recent research points to knowledge distillation as a potential solution, showing that when training data for a given task is abundant, it is possible to distill a large (teacher) LM into a small task-specific (student) network with minimal loss of performance. However, when such data is scarce, there remains a significant performance gap between large pretrained LMs and smaller task-specific models, even when training via distillation. In this paper, we bridge this gap with a novel training approach, called generation-distillation, that leverages large finetuned LMs in two ways: (1) to generate new (unlabeled) training examples, and (2) to distill their knowledge into a small network using these examples. Across three low-resource text classification datsets, we achieve comparable performance to BERT while using 300x fewer parameters, and we outperform prior approaches to distillation for text classification while using 3x fewer parameters.
Off-policy ensemble reinforcement learning (RL) methods have demonstrated impressive results across a range of RL benchmark tasks. Recent works suggest that directly imitating experts' policies in a supervised manner before or during the course of training enables faster policy improvement for an RL agent. Motivated by these recent insights, we propose Periodic Intra-Ensemble Knowledge Distillation (PIEKD). PIEKD is a learning framework that uses an ensemble of policies to act in the environment while periodically sharing knowledge amongst policies in the ensemble through knowledge distillation. Our experiments demonstrate that PIEKD improves upon a state-of-the-art RL method in sample efficiency on several challenging MuJoCo benchmark tasks. Additionally, we perform ablation studies to better understand PIEKD.
Knowledge Distillation, as a model compression technique, has received great attention. The knowledge of a well-performed teacher is distilled to a student with a small architecture. The architecture of the small student is often chosen to be similar to their teacher's, with fewer layers or fewer channels, or both. However, even with the same number of FLOPs or parameters, the students with different architecture can achieve different generalization ability. The configuration of a student architecture requires intensive network architecture engineering. In this work, instead of designing a good student architecture manually, we propose to search for the optimal student automatically. Based on L1-norm optimization, a subgraph from the teacher network topology graph is selected as a student, the goal of which is to minimize the KL-divergence between student's and teacher's outputs. We verify the proposal on CIFAR10 and CIFAR100 datasets. The empirical experiments show that the learned student architecture achieves better performance than ones specified manually. We also visualize and understand the architecture of the found student.
Deep neural networks achieve state-of-the-art performance for a range of classification and inference tasks. However, the use of stochastic gradient descent combined with the nonconvexity of the underlying optimization problems renders parameter learning susceptible to initialization. To address this issue, a variety of methods that rely on random parameter initialization or knowledge distillation have been proposed in the past. In this paper, we propose FuseInit, a novel method to initialize shallower networks by fusing neighboring layers of deeper networks that are trained with random initialization. We develop theoretical results and efficient algorithms for mean-square error (MSE)-optimal fusion of neighboring dense-dense, convolutional-dense, and convolutional-convolutional layers. We show experiments for a range of classification and regression datasets, which suggest that deeper neural networks are less sensitive to initialization and shallower networks can perform better (sometimes as well as their deeper counterparts) if initialized with FuseInit.
Practitioners often need to build ASR systems for new use cases in a short amount of time, given limited in-domain data. While recently developed end-to-end methods largely simplify the modeling pipelines, they still suffer from the data sparsity issue. In this work, we explore a few simple-to-implement techniques for building online ASR systems in an end-to-end fashion, with a small amount of transcribed data in the target domain. These techniques include data augmentation in the target domain, domain adaptation using models previously trained on a large source domain, and knowledge distillation on non-transcribed target domain data, using an adapted bi-directional model as the teacher; they are applicable in real scenarios with different types of resources. Our experiments demonstrate that each technique is independently useful in the improvement of the online ASR performance in the target domain.
With the explosive growth of online products and content, recommendation techniques have been considered as an effective tool to overcome information overload, improve user experience, and boost business revenue. In recent years, we have observed a new desideratum of considering long-term rewards of multiple related recommendation tasks simultaneously. The consideration of long-term rewards is strongly tied to business revenue and growth. Learning multiple tasks simultaneously could generally improve the performance of individual task due to knowledge sharing in multi-task learning. While a few existing works have studied long-term rewards in recommendations, they mainly focus on a single recommendation task. In this paper, we propose {\it PoDiRe}: a \underline{po}licy \underline{di}stilled \underline{re}commender that can address long-term rewards of recommendations and simultaneously handle multiple recommendation tasks. This novel recommendation solution is based on a marriage of deep reinforcement learning and knowledge distillation techniques, which is able to establish knowledge sharing among different tasks and reduce the size of a learning model. The resulting model is expected to attain better performance and lower response latency for real-time recommendation services. In collaboration with Samsung Game Launcher, one of the world's largest commercial mobile game platforms, we conduct a comprehensive experimental study on large-scale real data with hundreds of millions of events and show that our solution outperforms many state-of-the-art methods in terms of several standard evaluation metrics.
We present a filter pruning approach for deep model compression, using a multitask network. Our approach is based on learning a a pruner network to prune a pre-trained target network. The pruner is essentially a multitask deep neural network with binary outputs that help identify the filters from each layer of the original network that do not have any significant contribution to the model and can therefore be pruned. The pruner network has the same architecture as the original network except that it has a multitask/multi-output last layer containing binary-valued outputs (one per filter), which indicate which filters have to be pruned. The pruner's goal is to minimize the number of filters from the original network by assigning zero weights to the corresponding output feature-maps. In contrast to most of the existing methods, instead of relying on iterative pruning, our approach can prune the network (original network) in one go and, moreover, does not require specifying the degree of pruning for each layer (and can learn it instead). The compressed model produced by our approach is generic and does not need any special hardware/software support. Moreover, augmenting with other methods such as knowledge distillation, quantization, and connection pruning can increase the degree of compression for the proposed approach. We show the efficacy of our proposed approach for classification and object detection tasks.
We present MoVNect, a lightweight deep neural network to capture 3D human pose using a single RGB camera. To improve the overall performance of the model, we apply the teacher-student learning method based knowledge distillation to 3D human pose estimation. Real-time post-processing makes the CNN output yield temporally stable 3D skeletal information, which can be used in applications directly. We implement a 3D avatar application running on mobile in real-time to demonstrate that our network achieves both high accuracy and fast inference time. Extensive evaluations show the advantages of our lightweight model with the proposed training method over previous 3D pose estimation methods on the Human3.6M dataset and mobile devices.
Object re-identification (re-id) aims to identify a specific object across times or camera views, with the person re-id and vehicle re-id as the most widely studied applications. Re-id is challenging because of the variations in viewpoints, (human) poses, and occlusions. Multi-shots of the same object can cover diverse viewpoints/poses and thus provide more comprehensive information. In this paper, we propose exploiting the multi-shots of the same identity to guide the feature learning of each individual image. Specifically, we design an Uncertainty-aware Multi-shot Teacher-Student (UMTS) Network. It consists of a teacher network (T-net) that learns the comprehensive features from multiple images of the same object, and a student network (S-net) that takes a single image as input. In particular, we take into account the data dependent heteroscedastic uncertainty for effectively transferring the knowledge from the T-net to S-net. To the best of our knowledge, we are the first to make use of multi-shots of an object in a teacher-student learning manner for effectively boosting the single image based re-id. We validate the effectiveness of our approach on the popular vehicle re-id and person re-id datasets. In inference, the S-net alone significantly outperforms the baselines and achieves the state-of-the-art performance.
The success of deep learning has brought forth a wave of interest in computer hardware design to better meet the high demands of neural network inference. In particular, analog computing hardware has been heavily motivated specifically for accelerating neural networks, based on either electronic, optical or photonic devices, which may well achieve lower power consumption than conventional digital electronics. However, these proposed analog accelerators suffer from the intrinsic noise generated by their physical components, which makes it challenging to achieve high accuracy on deep neural networks. Hence, for successful deployment on analog accelerators, it is essential to be able to train deep neural networks to be robust to random continuous noise in the network weights, which is a somewhat new challenge in machine learning. In this paper, we advance the understanding of noisy neural networks. We outline how a noisy neural network has reduced learning capacity as a result of loss of mutual information between its input and output. To combat this, we propose using knowledge distillation combined with noise injection during training to achieve more noise robust networks, which is demonstrated experimentally across different networks and datasets, including ImageNet. Our method achieves models with as much as two times greater noise tolerance compared with the previous best attempts, which is a significant step towards making analog hardware practical for deep learning.
Large pre-trained language models such as BERT have shown their effectiveness in various natural language processing tasks. However, the huge parameter size makes them difficult to be deployed in real-time applications that require quick inference with limited resources. Existing methods compress BERT into small models while such compression is task-independent, i.e., the same compressed BERT for all different downstream tasks. Motivated by the necessity and benefits of task-oriented BERT compression, we propose a novel compression method, AdaBERT, that leverages differentiable Neural Architecture Search to automatically compress BERT into task-adaptive small models for specific tasks. We incorporate a task-oriented knowledge distillation loss to provide search hints and an efficiency-aware loss as search constraints, which enables a good trade-off between efficiency and effectiveness for task-adaptive BERT compression. We evaluate AdaBERT on several NLP tasks, and the results demonstrate that those task-adaptive compressed models are 12.7x to 29.3x faster than BERT in inference time and 11.5x to 17.0x smaller in terms of parameter size, while comparable performance is maintained.
Multi-modal learning is typically performed with network architectures containing modality-specific layers and shared layers, utilizing co-registered images of different modalities. We propose a novel learning scheme for unpaired cross-modality image segmentation, with a highly compact architecture achieving superior segmentation accuracy. In our method, we heavily reuse network parameters, by sharing all convolutional kernels across CT and MRI, and only employ modality-specific internal normalization layers which compute respective statistics. To effectively train such a highly compact model, we introduce a novel loss term inspired by knowledge distillation, by explicitly constraining the KL-divergence of our derived prediction distributions between modalities. We have extensively validated our approach on two multi-class segmentation problems: i) cardiac structure segmentation, and ii) abdominal organ segmentation. Different network settings, i.e., 2D dilated network and 3D U-net, are utilized to investigate our method's general efficacy. Experimental results on both tasks demonstrate that our novel multi-modal learning scheme consistently outperforms single-modal training and previous multi-modal approaches.
In real-world scenarios, data tends to exhibit a long-tailed distribution, which increases the difficulty of training deep networks. In this paper, we propose a novel self-paced knowledge distillation framework, termed Learning From Multiple Experts (LFME). Our method is inspired by the observation that networks trained on less imbalanced subsets of the distribution often yield better performances than their jointly-trained counterparts. We refer to these models as 'Experts', and the proposed LFME framework aggregates the knowledge from multiple 'Experts' to learn a unified student model. Specifically, the proposed framework involves two levels of adaptive learning schedules: Self-paced Expert Selection and Curriculum Instance Selection, so that the knowledge is adaptively transferred to the 'Student'. We conduct extensive experiments and demonstrate that our method is able to achieve superior performances compared to state-of-the-art methods. We also show that our method can be easily plugged into state-of-the-art long-tailed classification algorithms for further improvements.
Knowledge distillation (KD) is a new method for transferring knowledge of a structure under training to another one. The typical application of KD is in the form of learning a small model (named as a student) by soft labels produced by a complex model (named as a teacher). Due to the novel idea introduced in KD, recently, its notion is used in different methods such as compression and processes that are going to enhance the model accuracy. Although different techniques are proposed in the area of KD, there is a lack of a model to generalize KD techniques. In this paper, various studies in the scope of KD are investigated and analyzed to build a general model for KD. All the methods and techniques in KD can be summarized through the proposed model. By utilizing the proposed model, different methods in KD are better investigated and explored. The advantages and disadvantages of different approaches in KD can be better understood and develop a new strategy for KD can be possible. Using the proposed model, different KD methods are represented in an abstract view.