Zhe Li

Zhendong Yang, Zhe Li, Mingqi Shao et al.

Knowledge distillation has been applied to various tasks successfully. The current distillation algorithm usually improves students' performance by imitating the output of the teacher. This paper shows that teachers can also improve students' representation power by guiding students' feature recovery. From this point of view, we propose Masked Generative Distillation (MGD), which is simple: we mask random pixels of the student's feature and force it to generate the teacher's full feature through a simple block. MGD is a truly general feature-based distillation method, which can be utilized on various tasks, including image classification, object detection, semantic segmentation and instance segmentation. We experiment on different models with extensive datasets and the results show that all the students achieve excellent improvements. Notably, we boost ResNet-18 from 69.90% to 71.69% ImageNet top-1 accuracy, RetinaNet with ResNet-50 backbone from 37.4 to 41.0 Boundingbox mAP, SOLO based on ResNet-50 from 33.1 to 36.2 Mask mAP and DeepLabV3 based on ResNet-18 from 73.20 to 76.02 mIoU. Our codes are available at https://github.com/yzd-v/MGD.

Zhendong Yang, Ailing Zeng, Zhe Li et al.

Knowledge Distillation (KD) uses the teacher's prediction logits as soft labels to guide the student, while self-KD does not need a real teacher to require the soft labels. This work unifies the formulations of the two tasks by decomposing and reorganizing the generic KD loss into a Normalized KD (NKD) loss and customized soft labels for both target class (image's category) and non-target classes named Universal Self-Knowledge Distillation (USKD). We decompose the KD loss and find the non-target loss from it forces the student's non-target logits to match the teacher's, but the sum of the two non-target logits is different, preventing them from being identical. NKD normalizes the non-target logits to equalize their sum. It can be generally used for KD and self-KD to better use the soft labels for distillation loss. USKD generates customized soft labels for both target and non-target classes without a teacher. It smooths the target logit of the student as the soft target label and uses the rank of the intermediate feature to generate the soft non-target labels with Zipf's law. For KD with teachers, our NKD achieves state-of-the-art performance on CIFAR-100 and ImageNet datasets, boosting the ImageNet Top-1 accuracy of ResNet18 from 69.90% to 71.96% with a ResNet-34 teacher. For self-KD without teachers, USKD is the first self-KD method that can be effectively applied to both CNN and ViT models with negligible additional time and memory cost, resulting in new state-of-the-art results, such as 1.17% and 0.55% accuracy gains on ImageNet for MobileNet and DeiT-Tiny, respectively. Our codes are available at https://github.com/yzd-v/cls_KD.

Zhendong Yang, Zhe Li, Ailing Zeng et al.

Knowledge Distillation (KD) for Convolutional Neural Network (CNN) is extensively studied as a way to boost the performance of a small model. Recently, Vision Transformer (ViT) has achieved great success on many computer vision tasks and KD for ViT is also desired. However, besides the output logit-based KD, other feature-based KD methods for CNNs cannot be directly applied to ViT due to the huge structure gap. In this paper, we explore the way of feature-based distillation for ViT. Based on the nature of feature maps in ViT, we design a series of controlled experiments and derive three practical guidelines for ViT's feature distillation. Some of our findings are even opposite to the practices in the CNN era. Based on the three guidelines, we propose our feature-based method ViTKD which brings consistent and considerable improvement to the student. On ImageNet-1k, we boost DeiT-Tiny from 74.42% to 76.06%, DeiT-Small from 80.55% to 81.95%, and DeiT-Base from 81.76% to 83.46%. Moreover, ViTKD and the logit-based KD method are complementary and can be applied together directly. This combination can further improve the performance of the student. Specifically, the student DeiT-Tiny, Small, and Base achieve 77.78%, 83.59%, and 85.41%, respectively. The code is available at https://github.com/yzd-v/cls_KD.

Liang Jiang, Zujie Wen, Zhongping Liang et al.

In the past decade, there has been substantial progress at training increasingly deep neural networks. Recent advances within the teacher--student training paradigm have established that information about past training updates show promise as a source of guidance during subsequent training steps. Based on this notion, in this paper, we propose Long Short-Term Sample Distillation, a novel training policy that simultaneously leverages multiple phases of the previous training process to guide the later training updates to a neural network, while efficiently proceeding in just one single generation pass. With Long Short-Term Sample Distillation, the supervision signal for each sample is decomposed into two parts: a long-term signal and a short-term one. The long-term teacher draws on snapshots from several epochs ago in order to provide steadfast guidance and to guarantee teacher--student differences, while the short-term one yields more up-to-date cues with the goal of enabling higher-quality updates. Moreover, the teachers for each sample are unique, such that, overall, the model learns from a very diverse set of teachers. Comprehensive experimental results across a range of vision and NLP tasks demonstrate the effectiveness of this new training method.

Yuyin Zhou, Zhe Li, Song Bai et al.

Accurate multi-organ abdominal CT segmentation is essential to many clinical applications such as computer-aided intervention. As data annotation requires massive human labor from experienced radiologists, it is common that training data are partially labeled, e.g., pancreas datasets only have the pancreas labeled while leaving the rest marked as background. However, these background labels can be misleading in multi-organ segmentation since the "background" usually contains some other organs of interest. To address the background ambiguity in these partially-labeled datasets, we propose Prior-aware Neural Network (PaNN) via explicitly incorporating anatomical priors on abdominal organ sizes, guiding the training process with domain-specific knowledge. More specifically, PaNN assumes that the average organ size distributions in the abdomen should approximate their empirical distributions, a prior statistics obtained from the fully-labeled dataset. As our training objective is difficult to be directly optimized using stochastic gradient descent [20], we propose to reformulate it in a min-max form and optimize it via the stochastic primal-dual gradient algorithm. PaNN achieves state-of-the-art performance on the MICCAI2015 challenge "Multi-Atlas Labeling Beyond the Cranial Vault", a competition on organ segmentation in the abdomen. We report an average Dice score of 84.97%, surpassing the prior art by a large margin of 3.27%.

Siyu Liao, Zhe Li, Liang Zhao et al.

Deep neural networks (DNNs), especially deep convolutional neural networks (CNNs), have emerged as the powerful technique in various machine learning applications. However, the large model sizes of DNNs yield high demands on computation resource and weight storage, thereby limiting the practical deployment of DNNs. To overcome these limitations, this paper proposes to impose the circulant structure to the construction of convolutional layers, and hence leads to circulant convolutional layers (CircConvs) and circulant CNNs. The circulant structure and models can be either trained from scratch or re-trained from a pre-trained non-circulant model, thereby making it very flexible for different training environments. Through extensive experiments, such strong structure-imposing approach is proved to be able to substantially reduce the number of parameters of convolutional layers and enable significant saving of computational cost by using fast multiplication of the circulant tensor.

Zhe Li, Caiwen Ding, Siyue Wang et al.

Recurrent Neural Networks (RNNs) are becoming increasingly important for time series-related applications which require efficient and real-time implementations. The two major types are Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks. It is a challenging task to have real-time, efficient, and accurate hardware RNN implementations because of the high sensitivity to imprecision accumulation and the requirement of special activation function implementations. A key limitation of the prior works is the lack of a systematic design optimization framework of RNN model and hardware implementations, especially when the block size (or compression ratio) should be jointly optimized with RNN type, layer size, etc. In this paper, we adopt the block-circulant matrix-based framework, and present the Efficient RNN (E-RNN) framework for FPGA implementations of the Automatic Speech Recognition (ASR) application. The overall goal is to improve performance/energy efficiency under accuracy requirement. We use the alternating direction method of multipliers (ADMM) technique for more accurate block-circulant training, and present two design explorations providing guidance on block size and reducing RNN training trials. Based on the two observations, we decompose E-RNN in two phases: Phase I on determining RNN model to reduce computation and storage subject to accuracy requirement, and Phase II on hardware implementations given RNN model, including processing element design/optimization, quantization, activation implementation, etc. Experimental results on actual FPGA deployments show that E-RNN achieves a maximum energy efficiency improvement of 37.4$\times$ compared with ESE, and more than 2$\times$ compared with C-LSTM, under the same accuracy.

Jian Ren, Zhe Li, Jianchao Yang et al.

Designing the structure of neural networks is considered one of the most challenging tasks in deep learning, especially when there is few prior knowledge about the task domain. In this paper, we propose an Ecologically-Inspired GENetic (EIGEN) approach that uses the concept of succession, extinction, mimicry, and gene duplication to search neural network structure from scratch with poorly initialized simple network and few constraints forced during the evolution, as we assume no prior knowledge about the task domain. Specifically, we first use primary succession to rapidly evolve a population of poorly initialized neural network structures into a more diverse population, followed by a secondary succession stage for fine-grained searching based on the networks from the primary succession. Extinction is applied in both stages to reduce computational cost. Mimicry is employed during the entire evolution process to help the inferior networks imitate the behavior of a superior network and gene duplication is utilized to duplicate the learned blocks of novel structures, both of which help to find better network structures. Experimental results show that our proposed approach can achieve similar or better performance compared to the existing genetic approaches with dramatically reduced computation cost. For example, the network discovered by our approach on CIFAR-100 dataset achieves 78.1% test accuracy under 120 GPU hours, compared to 77.0% test accuracy in more than 65, 536 GPU hours in [35].

Zhe Li, Ji Li, Ao Ren et al.

Recently, Deep Convolutional Neural Network (DCNN) has achieved tremendous success in many machine learning applications. Nevertheless, the deep structure has brought significant increases in computation complexity. Largescale deep learning systems mainly operate in high-performance server clusters, thus restricting the application extensions to personal or mobile devices. Previous works on GPU and/or FPGA acceleration for DCNNs show increasing speedup, but ignore other constraints, such as area, power, and energy. Stochastic Computing (SC), as a unique data representation and processing technique, has the potential to enable the design of fully parallel and scalable hardware implementations of large-scale deep learning systems. This paper proposed an automatic design allocation algorithm driven by budget requirement considering overall accuracy performance. This systematic method enables the automatic design of a DCNN where all design parameters are jointly optimized. Experimental results demonstrate that proposed algorithm can achieve a joint optimization of all design parameters given the comprehensive budget of a DCNN.

Zhe Li, Shuo Wang, Caiwen Ding et al.

Recurrent Neural Networks (RNNs) are becoming increasingly important for time series-related applications which require efficient and real-time implementations. The recent pruning based work ESE suffers from degradation of performance/energy efficiency due to the irregular network structure after pruning. We propose block-circulant matrices for weight matrix representation in RNNs, thereby achieving simultaneous model compression and acceleration. We aim to implement RNNs in FPGA with highest performance and energy efficiency, with certain accuracy requirement (negligible accuracy degradation). Experimental results on actual FPGA deployments shows that the proposed framework achieves a maximum energy efficiency improvement of 35.7$\times$ compared with ESE.

Zhe Li, Xiaolong Ma, Hongjia Li et al.

Increasing malicious users have sought practices to leverage 3D printing technology to produce unlawful tools in criminal activities. Current regulations are inadequate to deal with the rapid growth of 3D printers. It is of vital importance to enable 3D printers to identify the objects to be printed, so that the manufacturing procedure of an illegal weapon can be terminated at the early stage. Deep learning yields significant rises in performance in the object recognition tasks. However, the lack of large-scale databases in 3D printing domain stalls the advancement of automatic illegal weapon recognition. This paper presents a new 3D printing image database, namely C3PO, which compromises two subsets for the different system working scenarios. We extract images from the numerical control programming code files of 22 3D models, and then categorize the images into 10 distinct labels. The first set consists of 62,200 images which represent the object projections on the three planes in a Cartesian coordinate system. And the second sets consists of sequences of total 671,677 images to simulate the cameras' captures of the printed objects. Importantly, we demonstrate that the weapons can be recognized in either scenario using deep learning based approaches using our proposed database. % We also use the trained deep models to build a prototype of object-aware 3D printer. The quantitative results are promising, and the future exploration of the database and the crime prevention in 3D printing are demanding tasks.

Xiaolong Ma, Yipeng Zhang, Geng Yuan et al.

With recent trend of wearable devices and Internet of Things (IoTs), it becomes attractive to develop hardware-based deep convolutional neural networks (DCNNs) for embedded applications, which require low power/energy consumptions and small hardware footprints. Recent works demonstrated that the Stochastic Computing (SC) technique can radically simplify the hardware implementation of arithmetic units and has the potential to satisfy the stringent power requirements in embedded devices. However, in these works, the memory design optimization is neglected for weight storage, which will inevitably result in large hardware cost. Moreover, if conventional volatile SRAM or DRAM cells are utilized for weight storage, the weights need to be re-initialized whenever the DCNN platform is re-started. In order to overcome these limitations, in this work we adopt an emerging non-volatile Domain-Wall Memory (DWM), which can achieve ultra-high density, to replace SRAM for weight storage in SC-based DCNNs. We propose DW-CNN, the first comprehensive design optimization framework of DWM-based weight storage method. We derive the optimal memory type, precision, and organization, as well as whether to store binary or stochastic numbers. We present effective resource sharing scheme for DWM-based weight storage in the convolutional and fully-connected layers of SC-based DCNNs to achieve a desirable balance among area, power (energy) consumption, and application-level accuracy.

Ao Ren, Ji Li, Zhe Li et al.

With recent advancing of Internet of Things (IoTs), it becomes very attractive to implement the deep convolutional neural networks (DCNNs) onto embedded/portable systems. Presently, executing the software-based DCNNs requires high-performance server clusters in practice, restricting their widespread deployment on the mobile devices. To overcome this issue, considerable research efforts have been conducted in the context of developing highly-parallel and specific DCNN hardware, utilizing GPGPUs, FPGAs, and ASICs. Stochastic Computing (SC), which uses bit-stream to represent a number within [-1, 1] by counting the number of ones in the bit-stream, has a high potential for implementing DCNNs with high scalability and ultra-low hardware footprint. Since multiplications and additions can be calculated using AND gates and multiplexers in SC, significant reductions in power/energy and hardware footprint can be achieved compared to the conventional binary arithmetic implementations. The tremendous savings in power (energy) and hardware resources bring about immense design space for enhancing scalability and robustness for hardware DCNNs. This paper presents the first comprehensive design and optimization framework of SC-based DCNNs (SC-DCNNs). We first present the optimal designs of function blocks that perform the basic operations, i.e., inner product, pooling, and activation function. Then we propose the optimal design of four types of combinations of basic function blocks, named feature extraction blocks, which are in charge of extracting features from input feature maps. Besides, weight storage methods are investigated to reduce the area and power/energy consumption for storing weights. Finally, the whole SC-DCNN implementation is optimized, with feature extraction blocks carefully selected, to minimize area and power/energy consumption while maintaining a high network accuracy level.

Zhe Li, Boqing Gong, Tianbao Yang

Dropout has been witnessed with great success in training deep neural networks by independently zeroing out the outputs of neurons at random. It has also received a surge of interest for shallow learning, e.g., logistic regression. However, the independent sampling for dropout could be suboptimal for the sake of convergence. In this paper, we propose to use multinomial sampling for dropout, i.e., sampling features or neurons according to a multinomial distribution with different probabilities for different features/neurons. To exhibit the optimal dropout probabilities, we analyze the shallow learning with multinomial dropout and establish the risk bound for stochastic optimization. By minimizing a sampling dependent factor in the risk bound, we obtain a distribution-dependent dropout with sampling probabilities dependent on the second order statistics of the data distribution. To tackle the issue of evolving distribution of neurons in deep learning, we propose an efficient adaptive dropout (named \textbf{evolutional dropout}) that computes the sampling probabilities on-the-fly from a mini-batch of examples. Empirical studies on several benchmark datasets demonstrate that the proposed dropouts achieve not only much faster convergence and but also a smaller testing error than the standard dropout. For example, on the CIFAR-100 data, the evolutional dropout achieves relative improvements over 10\% on the prediction performance and over 50\% on the convergence speed compared to the standard dropout.