Collaborative Teacher-Student Learning via Multiple Knowledge Transfer

Knowledge distillation (KD), as an efficient and effective model compression technique, has been receiving considerable attention in deep learning. The key to its success is to transfer knowledge from a large teacher network to a small student one. However, most of the existing knowledge distillation methods consider only one type of knowledge learned from either instance features or instance relations via a specific distillation strategy in teacher-student learning. There are few works that explore the idea of transferring different types of knowledge with different distillation strategies in a unified framework. Moreover, the frequently used offline distillation suffers from a limited learning capacity due to the fixed teacher-student architecture. In this paper we propose a collaborative teacher-student learning via multiple knowledge transfer (CTSL-MKT) that prompts both self-learning and collaborative learning. It allows multiple students learn knowledge from both individual instances and instance relations in a collaborative way. While learning from themselves with self-distillation, they can also guide each other via online distillation. The experiments and ablation studies on four image datasets demonstrate that the proposed CTSL-MKT significantly outperforms the state-of-the-art KD methods.

Understanding and Improving Lexical Choice in Non-Autoregressive Translation

Knowledge distillation (KD) is essential for training non-autoregressive translation (NAT) models by reducing the complexity of the raw data with an autoregressive teacher model. In this study, we empirically show that as a side effect of this training, the lexical choice errors on low-frequency words are propagated to the NAT model from the teacher model. To alleviate this problem, we propose to expose the raw data to NAT models to restore the useful information of low-frequency words, which are missed in the distilled data. To this end, we introduce an extra Kullback-Leibler divergence term derived by comparing the lexical choice of NAT model and that embedded in the raw data. Experimental results across language pairs and model architectures demonstrate the effectiveness and universality of the proposed approach. Extensive analyses confirm our claim that our approach improves performance by reducing the lexical choice errors on low-frequency words. Encouragingly, our approach pushes the SOTA NAT performance on the WMT14 English-German and WMT16 Romanian-English datasets up to 27.8 and 33.8 BLEU points, respectively. The source code will be released.

AdderNet and its Minimalist Hardware Design for Energy-Efficient Artificial Intelligence

Convolutional neural networks (CNN) have been widely used for boosting the performance of many machine intelligence tasks. However, the CNN models are usually computationally intensive and energy consuming, since they are often designed with numerous multiply-operations and considerable parameters for the accuracy reason. Thus, it is difficult to directly apply them in the resource-constrained environments such as 'Internet of Things' (IoT) devices and smart phones. To reduce the computational complexity and energy burden, here we present a novel minimalist hardware architecture using adder convolutional neural network (AdderNet), in which the original convolution is replaced by adder kernel using only additions. To maximally excavate the potential energy consumption, we explore the low-bit quantization algorithm for AdderNet with shared-scaling-factor method, and we design both specific and general-purpose hardware accelerators for AdderNet. Experimental results show that the adder kernel with int8/int16 quantization also exhibits high performance, meanwhile consuming much less resources (theoretically ~81% off). In addition, we deploy the quantized AdderNet on FPGA (Field Programmable Gate Array) platform. The whole AdderNet can practically achieve 16% enhancement in speed, 67.6%-71.4% decrease in logic resource utilization and 47.85%-77.9% decrease in power consumption compared to CNN under the same circuit architecture. With a comprehensive comparison on the performance, power consumption, hardware resource consumption and network generalization capability, we conclude the AdderNet is able to surpass all the other competitors including the classical CNN, novel memristor-network, XNOR-Net and the shift-kernel based network, indicating its great potential in future high performance and energy-efficient artificial intelligence applications.

CAA : Channelized Axial Attention for Semantic Segmentation

Self-attention and channel attention, modelling the semantic interdependencies in spatial and channel dimensions respectively, have recently been widely used for semantic segmentation. However, computing self-attention and channel attention separately and then fusing them directly can cause conflicting feature representations. In this paper, we propose the Channelized Axial Attention (CAA) to seamlessly integrate channel attention and axial attention with reduced computational complexity. After computing axial attention maps, we propose to channelize the intermediate results obtained from the transposed dot-product so that the channel importance of each axial representation is optimized across the whole receptive field. We further develop grouped vectorization, which allows our model to be run in the very limited GPU memory with a speed comparable with full vectorization. Comparative experiments conducted on multiple benchmark datasets, including Cityscapes, PASCAL Context and COCO-Stuff, demonstrate that our CAA not only requires much less computation resources but also outperforms the state-of-the-art segmentation models based on ResNet-101 on all tested datasets.

Bayesian Inference Forgetting

The right to be forgotten has been legislated in many countries but the enforcement in machine learning would cause unbearable costs: companies may need to delete whole models trained from massive resources because of single individual requests. Existing works propose to remove the influence of the requested datums on the learned models via its influence function which is no longer naturally well-defined in Bayesian inference. To address this problem, this paper proposes a {\it Bayesian inference forgetting} (BIF) framework to extend the applicable domain to Bayesian inference. In the BIF framework, we develop forgetting algorithms for variational inference and Markov chain Monte Carlo. We show that our algorithms can provably remove the influence of single datums on the learned models. Theoretical analysis demonstrates that our algorithms have guaranteed generalizability. Experiments of Gaussian mixture models on the synthetic dataset and Bayesian neural networks on the Fashion-MNIST dataset verify the feasibility of our methods. The source code package is available at \url{https://github.com/fshp971/BIF}.

A Survey on Visual Transformer

Transformer is a type of deep neural network mainly based on self-attention mechanism which is originally applied in natural language processing field. Inspired by the strong representation ability of transformer, researchers propose to extend transformer for computer vision tasks. Transformer-based models show competitive and even better performance on various visual benchmarks compared to other network types such as convolutional networks and recurrent networks. With high performance and without inductive bias defined by human, transformer is receiving more and more attention from the visual community. In this paper we provide a literature review of these visual transformer models by categorizing them in different tasks and analyze the advantages and disadvantages of these methods. In particular, the main categories include the basic image classification, high-level vision, low-level vision and video processing. The self-attention in computer vision is also briefly revisited as self-attention is the base component in transformer. Efficient transformer methods are included for pushing transformer into real applications on the devices. Finally, we give a discussion about the challenges and further research directions for visual transformers.

Neural networks behave as hash encoders: An empirical study

The input space of a neural network with ReLU-like activations is partitioned into multiple linear regions, each corresponding to a specific activation pattern of the included ReLU-like activations. We demonstrate that this partition exhibits the following encoding properties across a variety of deep learning models: (1) {\it determinism}: almost every linear region contains at most one training example. We can therefore represent almost every training example by a unique activation pattern, which is parameterized by a {\it neural code}; and (2) {\it categorization}: according to the neural code, simple algorithms, such as $K$-Means, $K$-NN, and logistic regression, can achieve fairly good performance on both training and test data. These encoding properties surprisingly suggest that {\it normal neural networks well-trained for classification behave as hash encoders without any extra efforts.} In addition, the encoding properties exhibit variability in different scenarios. {Further experiments demonstrate that {\it model size}, {\it training time}, {\it training sample size}, {\it regularization}, and {\it label noise} contribute in shaping the encoding properties, while the impacts of the first three are dominant.} We then define an {\it activation hash phase chart} to represent the space expanded by {model size}, training time, training sample size, and the encoding properties, which is divided into three canonical regions: {\it under-expressive regime}, {\it critically-expressive regime}, and {\it sufficiently-expressive regime}. The source code package is available at \url{https://github.com/LeavesLei/activation-code}.

SPAGAN: Shortest Path Graph Attention Network

Graph convolutional networks (GCN) have recently demonstrated their potential in analyzing non-grid structure data that can be represented as graphs. The core idea is to encode the local topology of a graph, via convolutions, into the feature of a center node. In this paper, we propose a novel GCN model, which we term as Shortest Path Graph Attention Network (SPAGAN). Unlike conventional GCN models that carry out node-based attentions within each layer, the proposed SPAGAN conducts path-based attention that explicitly accounts for the influence of a sequence of nodes yielding the minimum cost, or shortest path, between the center node and its higher-order neighbors. SPAGAN therefore allows for a more informative and intact exploration of the graph structure and further {a} more effective aggregation of information from distant neighbors into the center node, as compared to node-based GCN methods. We test SPAGAN on the downstream classification task on several standard datasets, and achieve performances superior to the state of the art. Code is publicly available at https://github.com/ihollywhy/SPAGAN.