Search

Transformer-based pre-training models like BERT have achieved remarkable performance in many natural language processing tasks.However, these models are both computation and memory expensive, hindering their deployment to resource-constrained devices. In this work, we propose TernaryBERT, which ternarizes the weights in a fine-tuned BERT model. Specifically, we use both approximation-based and loss-aware ternarization methods and empirically investigate the ternarization granularity of different parts of BERT. Moreover, to reduce the accuracy degradation caused by the lower capacity of low bits, we leverage the knowledge distillation technique in the training process. Experiments on the GLUE benchmark and SQuAD show that our proposed TernaryBERT outperforms the other BERT quantization methods, and even achieves comparable performance as the full-precision model while being 14.9x smaller., Comment: Accepted by EMNLP 2020

The pre-trained language models like BERT, though powerful in many natural language processing tasks, are both computation and memory expensive. To alleviate this problem, one approach is to compress them for specific tasks before deployment. However, recent works on BERT compression usually compress the large BERT model to a fixed smaller size. They can not fully satisfy the requirements of different edge devices with various hardware performances. In this paper, we propose a novel dynamic BERT model (abbreviated as DynaBERT), which can flexibly adjust the size and latency by selecting adaptive width and depth. The training process of DynaBERT includes first training a width-adaptive BERT and then allowing both adaptive width and depth, by distilling knowledge from the full-sized model to small sub-networks. Network rewiring is also used to keep the more important attention heads and neurons shared by more sub-networks. Comprehensive experiments under various efficiency constraints demonstrate that our proposed dynamic BERT (or RoBERTa) at its largest size has comparable performance as BERT-base (or RoBERTa-base), while at smaller widths and depths consistently outperforms existing BERT compression methods. Code is available at https://github.com/huawei-noah/Pretrained-Language-Model/tree/master/DynaBERT., Comment: NeurIPS-2020 (Spotlight)

Long Range (LoRa) has become one of the most popular Low Power Wide Area (LPWA) technologies, which provides a desirable trade-off among communication range, battery life, and deployment cost. In LoRa networks, several transmission parameters can be allocated to ensure efficient and reliable communication. For example, the configuration of the spreading factor allows tuning the data rate and the transmission distance. However, how to dynamically adjust the setting that minimizes the collision probability while meeting the required communication performance is an open challenge. This paper proposes a novel Data Rate and Channel Control (DRCC) scheme for LoRa networks so as to improve wireless resource utilization and support a massive number of LoRa nodes. The scheme estimates channel conditions based on the short-term Data Extraction Rate (DER), and opportunistically adjusts the spreading factor to adapt the variation of channel conditions. Furthermore, the channel control is carried out to balance the link load of all available channels with the global information of the channel usage, which is able to lower the access collisions under dense deployments. Our experiments demonstrate that the proposed DRCC performs well on improving the reliability and capacity compared with other spreading factor allocation schemes in dense deployment scenarios., Comment: 6 pages, 6 figures

Long Range (LoRa) network is emerging as one of the most promising Low Power Wide Area (LPWA) networks, since it enables the energy-constraint devices distributed over wide areas to establish affordable connectivity. However, how to implement a cost-effective and flexible LoRa network is still an open challenge. This paper aims at exposing a feasible solution of design and implementation, allowing users to conveniently build a private LoRa network for various IoT applications. Firstly, several typical application scenarios of LoRa network are discussed. Then, the LoRa system architecture is presented with the functionality of each component. We address the hardware design and implementation of LoRa Gateway, which is the bridge between LoRa nodes and LoRa network server. Especially, the paper contributes by proposing an improved software architecture of LoRa network server whose source codes are open on GitHub. Under the architecture, LoRa network server is divided into four decoupled modules and uses the messaging system based on streaming data for the interaction between modules to guarantee scalability and flexibility. Finally, extensive experiments are conducted to evaluate the performance of LoRa networks in typical environments., Comment: Published version

The power law has been observed in the degree distributions of many biological neural networks. Sparse deep neural networks, which learn an economical representation from the data, resemble biological neural networks in many ways. In this paper, we study if these artificial networks also exhibit properties of the power law. Experimental results on two popular deep learning models, namely, multilayer perceptrons and convolutional neural networks, are affirmative. The power law is also naturally related to preferential attachment. To study the dynamical properties of deep networks in continual learning, we propose an internal preferential attachment model to explain how the network topology evolves. Experimental results show that with the arrival of a new task, the new connections made follow this preferential attachment process.

The huge size of deep networks hinders their use in small computing devices. In this paper, we consider compressing the network by weight quantization. We extend a recently proposed loss-aware weight binarization scheme to ternarization, with possibly different scaling parameters for the positive and negative weights, and m-bit (where m > 2) quantization. Experiments on feedforward and recurrent neural networks show that the proposed scheme outperforms state-of-the-art weight quantization algorithms, and is as accurate (or even more accurate) than the full-precision network.

Deep neural network models, though very powerful and highly successful, are computationally expensive in terms of space and time. Recently, there have been a number of attempts on binarizing the network weights and activations. This greatly reduces the network size, and replaces the underlying multiplications to additions or even XNOR bit operations. However, existing binarization schemes are based on simple matrix approximation and ignore the effect of binarization on the loss. In this paper, we propose a proximal Newton algorithm with diagonal Hessian approximation that directly minimizes the loss w.r.t. the binarized weights. The underlying proximal step has an efficient closed-form solution, and the second-order information can be efficiently obtained from the second moments already computed by the Adam optimizer. Experiments on both feedforward and recurrent networks show that the proposed loss-aware binarization algorithm outperforms existing binarization schemes, and is also more robust for wide and deep networks.

The Internet of Things (IoT), which enables common objects to be intelligent and interactive, is considered the next evolution of the Internet. Its pervasiveness and abilities to collect and analyze data which can be converted into information have motivated a plethora of IoT applications. For the successful deployment and management of these applications, cloud computing techniques are indispensable since they provide high computational capabilities as well as large storage capacity. This paper aims at providing insights about the architecture, implementation and performance of the IoT cloud. Several potential application scenarios of IoT cloud are studied, and an architecture is discussed regarding the functionality of each component. Moreover, the implementation details of the IoT cloud are presented along with the services that it offers. The main contributions of this paper lie in the combination of the Hypertext Transfer Protocol (HTTP) and Message Queuing Telemetry Transport (MQTT) servers to offer IoT services in the architecture of the IoT cloud with various techniques to guarantee high performance. Finally, experimental results are given in order to demonstrate the service capabilities of the IoT cloud under certain conditions., Comment: 19pages, 4figures, IEEE Communications Magazine

Autonomous driving technology has been regarded as a promising solution to reduce road accidents and traffic congestion, as well as to optimize the usage of fuel and lane. Reliable and high efficient Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure (V2I) communications are essential to let commercial autonomous driving vehicles be on the road before 2020. The current paper firstly presents the concept of Heterogeneous Vehicular NETworks (HetVNETs) for autonomous driving, in which an improved protocol stack is proposed to satisfy the communication requirements of not only safety but also non-safety services. We then consider and study in detail several typical scenarios for autonomous driving. In order to tackle the potential challenges raised by the autonomous driving vehicles in HetVNETs, new techniques from transmission to networking are proposed as potential solutions.

Heterogeneous Vehicular NETworks (HetVNETs) can meet various quality-of-service (QoS) requirements for intelligent transport system (ITS) services by integrating different access networks coherently. However, the current network architecture for HetVNET cannot efficiently deal with the increasing demands of rapidly changing network landscape. Thanks to the centralization and flexibility of the cloud radio access network (Cloud-RAN), soft-defined networking (SDN) can conveniently be applied to support the dynamic nature of future HetVNET functions and various applications while reducing the operating costs. In this paper, we first propose the multi-layer Cloud RAN architecture for implementing the new network, where the multi-domain resources can be exploited as needed for vehicle users. Then, the high-level design of soft-defined HetVNET is presented in detail. Finally, we briefly discuss key challenges and solutions for this new network, corroborating its feasibility in the emerging fifth-generation (5G) era.