The highest accuracy object detectors to date are based on a two-stage approach popularized by R-CNN, where a classifier is applied to a sparse set of candidate object locations. In contrast, one-stage detectors that are applied over a regular, dense sampling of possible object locations have the potential to be faster and simpler, but have trailed the accuracy of two-stage detectors thus far. In this paper, we investigate why this is the case. We discover that the extreme foreground-background class imbalance encountered during training of dense detectors is the central cause. We propose to address this class imbalance by reshaping the standard cross entropy loss such that it down-weights the loss assigned to well-classified examples. Our novel Focal Loss focuses training on a sparse set of hard examples and prevents the vast number of easy negatives from overwhelming the detector during training. To evaluate the effectiveness of our loss, we design and train a simple dense detector we call RetinaNet. Our results show that when trained with the focal loss, RetinaNet is able to match the speed of previous one-stage detectors while surpassing the accuracy of all existing state-of-the-art two-stage detectors. Code is at: https://github.com/facebookresearch/Detectron.
In this paper, we propose a simple but effective method for training neural networks with a limited amount of training data. Our approach inherits the idea of knowledge distillation that transfers knowledge from a deep or wide reference model to a shallow or narrow target model. The proposed method employs this idea to mimic predictions of reference estimators that are more robust against overfitting than the network we want to train. Different from almost all the previous work for knowledge distillation that requires a large amount of labeled training data, the proposed method requires only a small amount of training data. Instead, we introduce pseudo training examples that are optimized as a part of model parameters. Experimental results for several benchmark datasets demonstrate that the proposed method outperformed all the other baselines, such as naive training of the target model and standard knowledge distillation.
Existing approaches to neural machine translation condition each output word on previously generated outputs. We introduce a model that avoids this autoregressive property and produces its outputs in parallel, allowing an order of magnitude lower latency during inference. Through knowledge distillation, the use of input token fertilities as a latent variable, and policy gradient fine-tuning, we achieve this at a cost of as little as 2.0 BLEU points relative to the autoregressive Transformer network used as a teacher. We demonstrate substantial cumulative improvements associated with each of the three aspects of our training strategy, and validate our approach on IWSLT 2016 English-German and two WMT language pairs. By sampling fertilities in parallel at inference time, our non-autoregressive model achieves near-state-of-the-art performance of 29.8 BLEU on WMT 2016 English-Romanian.
Despite their prevalence, deep networks are poorly understood. This is due, at least in part, to their highly parameterized nature. As such, while certain structures have been found to work better than others, the significance of a model's unique structure, or the importance of a given layer, and how these translate to overall accuracy, remains unclear. In this paper, we analyze these properties of deep neural networks via a process we term deep net triage. Like medical triage---the assessment of the importance of various wounds---we assess the importance of layers in a neural network, or as we call it, their criticality. We do this by applying structural compression, whereby we reduce a block of layers to a single layer. After compressing a set of layers, we apply a combination of initialization and training schemes, and look at network accuracy, convergence, and the layer's learned filters to assess the criticality of the layer. We apply this analysis across four data sets of varying complexity. We find that the accuracy of the model does not depend on which layer was compressed; that accuracy can be recovered or exceeded after compression by fine-tuning across the entire model; and, lastly, that Knowledge Distillation can be used to hasten convergence of a compressed network, but constrains the accuracy attainable to that of the base model.
Recently, there have been increasing demands to construct compact deep architectures to remove unnecessary redundancy and to improve the inference speed. While many recent works focus on reducing the redundancy by eliminating unneeded weight parameters, it is not possible to apply a single deep architecture for multiple devices with different resources. When a new device or circumstantial condition requires a new deep architecture, it is necessary to construct and train a new network from scratch. In this work, we propose a novel deep learning framework, called a nested sparse network, which exploits an n-in-1-type nested structure in a neural network. A nested sparse network consists of multiple levels of networks with a different sparsity ratio associated with each level, and higher level networks share parameters with lower level networks to enable stable nested learning. The proposed framework realizes a resource-aware versatile architecture as the same network can meet diverse resource requirements. Moreover, the proposed nested network can learn different forms of knowledge in its internal networks at different levels, enabling multiple tasks using a single network, such as coarse-to-fine hierarchical classification. In order to train the proposed nested sparse network, we propose efficient weight connection learning and channel and layer scheduling strategies. We evaluate our network in multiple tasks, including adaptive deep compression, knowledge distillation, and learning class hierarchy, and demonstrate that nested sparse networks perform competitively, but more efficiently, compared to existing methods.
Recently, the booming fashion sector and its huge potential benefits have attracted tremendous attention from many research communities. In particular, increasing research efforts have been dedicated to the complementary clothing matching as matching clothes to make a suitable outfit has become a daily headache for many people, especially those who do not have the sense of aesthetics. Thanks to the remarkable success of neural networks in various applications such as image classification and speech recognition, the researchers are enabled to adopt the data-driven learning methods to analyze fashion items. Nevertheless, existing studies overlook the rich valuable knowledge (rules) accumulated in fashion domain, especially the rules regarding clothing matching. Towards this end, in this work, we shed light on complementary clothing matching by integrating the advanced deep neural networks and the rich fashion domain knowledge. Considering that the rules can be fuzzy and different rules may have different confidence levels to different samples, we present a neural compatibility modeling scheme with attentive knowledge distillation based on the teacher-student network scheme. Extensive experiments on the real-world dataset show the superiority of our model over several state-of-the-art baselines. Based upon the comparisons, we observe certain fashion insights that add value to the fashion matching study. As a byproduct, we released the codes, and involved parameters to benefit other researchers.
There is an increasing interest on accelerating neural networks for real-time applications. We study the student-teacher strategy, in which a small and fast student network is trained with the auxiliary information learned from a large and accurate teacher network. We propose to use conditional adversarial networks to learn the loss function to transfer knowledge from teacher to student. The proposed method is particularly effective for relatively small student networks. Moreover, experimental results show the effect of network size when the modern networks are used as student. We empirically study the trade-off between inference time and classification accuracy, and provide suggestions on choosing a proper student network.
Intelligent personal assistant systems with either text-based or voice-based conversational interfaces are becoming increasingly popular around the world. Retrieval-based conversation models have the advantages of returning fluent and informative responses. Most existing studies in this area are on open domain "chit-chat" conversations or task / transaction oriented conversations. More research is needed for information-seeking conversations. There is also a lack of modeling external knowledge beyond the dialog utterances among current conversational models. In this paper, we propose a learning framework on the top of deep neural matching networks that leverages external knowledge for response ranking in information-seeking conversation systems. We incorporate external knowledge into deep neural models with pseudo-relevance feedback and QA correspondence knowledge distillation. Extensive experiments with three information-seeking conversation data sets including both open benchmarks and commercial data show that, our methods outperform various baseline methods including several deep text matching models and the state-of-the-art method on response selection in multi-turn conversations. We also perform analysis over different response types, model variations and ranking examples. Our models and research findings provide new insights on how to utilize external knowledge with deep neural models for response selection and have implications for the design of the next generation of information-seeking conversation systems.
Over the last decade, Convolutional Neural Network (CNN) models have been highly successful in solving complex vision problems. However, these deep models are perceived as "black box" methods considering the lack of understanding of their internal functioning. There has been a significant recent interest in developing explainable deep learning models, and this paper is an effort in this direction. Building on a recently proposed method called Grad-CAM, we propose a generalized method called Grad-CAM++ that can provide better visual explanations of CNN model predictions, in terms of better object localization as well as explaining occurrences of multiple object instances in a single image, when compared to state-of-the-art. We provide a mathematical derivation for the proposed method, which uses a weighted combination of the positive partial derivatives of the last convolutional layer feature maps with respect to a specific class score as weights to generate a visual explanation for the corresponding class label. Our extensive experiments and evaluations, both subjective and objective, on standard datasets showed that Grad-CAM++ provides promising human-interpretable visual explanations for a given CNN architecture across multiple tasks including classification, image caption generation and 3D action recognition; as well as in new settings such as knowledge distillation.
Knowledge distillation (KD) consists of transferring knowledge from one machine learning model (the teacher}) to another (the student). Commonly, the teacher is a high-capacity model with formidable performance, while the student is more compact. By transferring knowledge, one hopes to benefit from the student's compactness. %we desire a compact model with performance close to the teacher's. We study KD from a new perspective: rather than compressing models, we train students parameterized identically to their teachers. Surprisingly, these {Born-Again Networks (BANs), outperform their teachers significantly, both on computer vision and language modeling tasks. Our experiments with BANs based on DenseNets demonstrate state-of-the-art performance on the CIFAR-10 (3.5%) and CIFAR-100 (15.5%) datasets, by validation error. Additional experiments explore two distillation objectives: (i) Confidence-Weighted by Teacher Max (CWTM) and (ii) Dark Knowledge with Permuted Predictions (DKPP). Both methods elucidate the essential components of KD, demonstrating a role of the teacher outputs on both predicted and non-predicted classes. We present experiments with students of various capacities, focusing on the under-explored case where students overpower teachers. Our experiments show significant advantages from transferring knowledge between DenseNets and ResNets in either direction.
We focus on the problem of training a deep neural network in generations. The flowchart is that, in order to optimize the target network (student), another network (teacher) with the same architecture is first trained, and used to provide part of supervision signals in the next stage. While this strategy leads to a higher accuracy, many aspects (e.g., why teacher-student optimization helps) still need further explorations. This paper studies this problem from a perspective of controlling the strictness in training the teacher network. Existing approaches mostly used a hard distribution (e.g., one-hot vectors) in training, leading to a strict teacher which itself has a high accuracy, but we argue that the teacher needs to be more tolerant, although this often implies a lower accuracy. The implementation is very easy, with merely an extra loss term added to the teacher network, facilitating a few secondary classes to emerge and complement to the primary class. Consequently, the teacher provides a milder supervision signal (a less peaked distribution), and makes it possible for the student to learn from inter-class similarity and potentially lower the risk of over-fitting. Experiments are performed on standard image classification tasks (CIFAR100 and ILSVRC2012). Although the teacher network behaves less powerful, the students show a persistent ability growth and eventually achieve higher classification accuracies than other competitors. Model ensemble and transfer feature extraction also verify the effectiveness of our approach.
Knowledge distillation compacts deep networks by letting a small student network learn from a large teacher network. The accuracy of knowledge distillation recently benefited from adding residual layers. We propose to reduce the size of the student network even further by recasting multiple residual layers in the teacher network into a single recurrent student layer. We propose three variants of adding recurrent connections into the student network, and show experimentally on CIFAR-10, Scenes and MiniPlaces, that we can reduce the number of parameters at little loss in accuracy.
Many recent works on knowledge distillation have provided ways to transfer the knowledge of a trained network for improving the learning process of a new one, but finding a good technique for knowledge distillation is still an open problem. In this paper, we provide a new perspective based on a decision boundary, which is one of the most important component of a classifier. The generalization performance of a classifier is closely related to the adequacy of its decision boundary, so a good classifier bears a good decision boundary. Therefore, transferring information closely related to the decision boundary can be a good attempt for knowledge distillation. To realize this goal, we utilize an adversarial attack to discover samples supporting a decision boundary. Based on this idea, to transfer more accurate information about the decision boundary, the proposed algorithm trains a student classifier based on the adversarial samples supporting the decision boundary. Experiments show that the proposed method indeed improves knowledge distillation and achieves the state-of-the-arts performance.
A thorough comprehension of image content demands a complex grasp of the interactions that may occur in the natural world. One of the key issues is to describe the visual relationships between objects. When dealing with real world data, capturing these very diverse interactions is a difficult problem. It can be alleviated by incorporating common sense in a network. For this, we propose a framework that makes use of semantic knowledge and estimates the relevance of object pairs during both training and test phases. Extracted from precomputed models and training annotations, this information is distilled into the neural network dedicated to this task. Using this approach, we observe a significant improvement on all classes of Visual Genome, a challenging visual relationship dataset. A 68.5% relative gain on the recall at 100 is directly related to the relevance estimate and a 32.7% gain to the knowledge distillation.
Deep neural networks with discrete latent variables offer the promise of better symbolic reasoning, and learning abstractions that are more useful to new tasks. There has been a surge in interest in discrete latent variable models, however, despite several recent improvements, the training of discrete latent variable models has remained challenging and their performance has mostly failed to match their continuous counterparts. Recent work on vector quantized autoencoders (VQ-VAE) has made substantial progress in this direction, with its perplexity almost matching that of a VAE on datasets such as CIFAR-10. In this work, we investigate an alternate training technique for VQ-VAE, inspired by its connection to the Expectation Maximization (EM) algorithm. Training the discrete bottleneck with EM helps us achieve better image generation results on CIFAR-10, and together with knowledge distillation, allows us to develop a non-autoregressive machine translation model whose accuracy almost matches a strong greedy autoregressive baseline Transformer, while being 3.3 times faster at inference.
In recent years, deep neural networks have achieved great success in the field of computer vision. However, it is still a big challenge to deploy these deep models on resource-constrained embedded devices such as mobile robots, smart phones and so on. Therefore, network compression for such platforms is a reasonable solution to reduce memory consumption and computation complexity. In this paper, a novel channel pruning method based on genetic algorithm is proposed to compress very deep Convolution Neural Networks (CNNs). Firstly, a pre-trained CNN model is pruned layer by layer according to the sensitivity of each layer. After that, the pruned model is fine-tuned based on knowledge distillation framework. These two improvements significantly decrease the model redundancy with less accuracy drop. Channel selection is a combinatorial optimization problem that has exponential solution space. In order to accelerate the selection process, the proposed method formulates it as a search problem, which can be solved efficiently by genetic algorithm. Meanwhile, a two-step approximation fitness function is designed to further improve the efficiency of genetic process. The proposed method has been verified on three benchmark datasets with two popular CNN models: VGGNet and ResNet. On the CIFAR-100 and ImageNet datasets, our approach outperforms several state-of-the-art methods. On the CIFAR-10 and SVHN datasets, the pruned VGGNet achieves better performance than the original model with 8 times parameters compression and 3 times FLOPs reduction.
This paper introduces channel gating, a dynamic, fine-grained, and hardware-efficient pruning scheme to reduce the computation cost for convolutional neural networks (CNNs). Channel gating identifies regions in the features that contribute less to the classification result, and skips the computation on a subset of the input channels for these ineffective regions. Unlike static network pruning, channel gating optimizes CNN inference at run-time by exploiting input-specific characteristics, which allows substantially reducing the compute cost with almost no accuracy loss. We experimentally show that applying channel gating in state-of-the-art networks achieves 2.7-8.0$\times$ reduction in floating-point operations (FLOPs) and 2.0-4.4$\times$ reduction in off-chip memory accesses with a minimal accuracy loss on CIFAR-10. Combining our method with knowledge distillation reduces the compute cost of ResNet-18 by 2.6$\times$ without accuracy drop on ImageNet. We further demonstrate that channel gating can be realized in hardware efficiently. Our approach exhibits sparsity patterns that are well-suited to dense systolic arrays with minimal additional hardware. We have designed an accelerator for channel gating networks, which can be implemented using either FPGAs or ASICs. Running a quantized ResNet-18 model for ImageNet, our accelerator achieves an encouraging speedup of 2.4$\times$ on average, with a theoretical FLOP reduction of 2.8$\times$.
We introduce collaborative learning in which multiple classifier heads of the same network are simultaneously trained on the same training data to improve generalization and robustness to label noise with no extra inference cost. It acquires the strengths from auxiliary training, multi-task learning and knowledge distillation. There are two important mechanisms involved in collaborative learning. First, the consensus of multiple views from different classifier heads on the same example provides supplementary information as well as regularization to each classifier, thereby improving generalization. Second, intermediate-level representation (ILR) sharing with backpropagation rescaling aggregates the gradient flows from all heads, which not only reduces training computational complexity, but also facilitates supervision to the shared layers. The empirical results on CIFAR and ImageNet datasets demonstrate that deep neural networks learned as a group in a collaborative way significantly reduce the generalization error and increase the robustness to label noise.
A large fraction of the arithmetic operations required to evaluate deep neural networks (DNNs) consists of matrix multiplications, in both convolution and fully connected layers. We perform end-to-end learning of low-cost approximations of matrix multiplications in DNN layers by casting matrix multiplications as 2-layer sum-product networks (SPNs) (arithmetic circuits) and learning their (ternary) edge weights from data. The SPNs disentangle multiplication and addition operations and enable us to impose a budget on the number of multiplication operations. Combining our method with knowledge distillation and applying it to image classification DNNs (trained on ImageNet) and language modeling DNNs (using LSTMs), we obtain a first-of-a-kind reduction in number of multiplications (over 99.5%) while maintaining the predictive performance of the full-precision models. Finally, we demonstrate that the proposed framework is able to rediscover Strassen's matrix multiplication algorithm, learning to multiply $2 \times 2$ matrices using only 7 multiplications instead of 8.
Knowledge distillation is effective to train small and generalisable network models for meeting the low-memory and fast running requirements. Existing offline distillation methods rely on a strong pre-trained teacher, which enables favourable knowledge discovery and transfer but requires a complex two-phase training procedure. Online counterparts address this limitation at the price of lacking a highcapacity teacher. In this work, we present an On-the-fly Native Ensemble (ONE) strategy for one-stage online distillation. Specifically, ONE trains only a single multi-branch network while simultaneously establishing a strong teacher on-the- fly to enhance the learning of target network. Extensive evaluations show that ONE improves the generalisation performance a variety of deep neural networks more significantly than alternative methods on four image classification dataset: CIFAR10, CIFAR100, SVHN, and ImageNet, whilst having the computational efficiency advantages.