- CCD and CMOS Imaging Sensors
- Advanced Neural Network Applications
- Atmospheric aerosols and clouds
- Advanced Memory and Neural Computing
- Atmospheric chemistry and aerosols
- Aeolian processes and effects
- Ferroelectric and Negative Capacitance Devices
- Neural Networks and Applications
- Parallel Computing and Optimization Techniques
- Meteorological Phenomena and Simulations
- Wireless Communication Networks Research
- Gait Recognition and Analysis
- Machine Fault Diagnosis Techniques
- Topic Modeling
- Precipitation Measurement and Analysis
- Transportation Planning and Optimization
- Blind Source Separation Techniques
- Wind and Air Flow Studies
- Indoor and Outdoor Localization Technologies
- Advanced Image and Video Retrieval Techniques
- Soil Moisture and Remote Sensing
- Speech and dialogue systems
- Security and Verification in Computing
- Traffic control and management
- Atmospheric Ozone and Climate
Fudan University
2000-2024
University of Hong Kong
2020-2024
Hefei University of Technology
2024
University of Science and Technology of China
2019-2023
New York University
2021
Chinese University of Hong Kong
2020
Zhongnan University of Economics and Law
2020
Abstract With the rapid growth of medical technology, intelligent healthcare has gradually integrated into our lives, and its core technology is human body motion state detection. According to basic tools being used, we divide this two categories: vision-based detection sensor-based detection, take fall monitoring as an example for elaboration. Finally, different methods are compared analyzed, difficulties that need be solved in field summarized, future trends included too.
To accelerate the inference of deep neural networks (DNNs), quantization with low-bitwidth numbers is actively researched. A prominent challenge to quantize DNN models into without significant accuracy degradation, especially at very low bitwidths (< 8 bits). This work targets an adaptive data representation variablelength encoding called DyBit. DyBit can dynamically adjust precision and range separate bit-fields be adapted weights/activations distribution. We also propose a hardware-aware...
Recurrent neural networks (RNNs) have been widely adopted in temporal sequence analysis, where realtime performance is often demand. However, RNNs suffer from heavy computational workload as the model comes with large weight matrices. Pruning (a compression method) schemes proposed for to eliminate redundant (close-to-zero) values. On one hand, non-structured pruning methods achieve a high rate but introducing computation irregularity (random sparsity), which unfriendly parallel hardware....
The cloud amount, referred to as the frequency of occurrences, is great importance for Earth–atmosphere system. It was conventionally quantified area fraction clouds in a given region, discarding three-dimensional nature both entities and their spatial distribution. Although explicit, it volume that fully depicts just related projection fraction. In this study, by using spaceborne radar measurements, distribution throughout troposphere investigated, contributions various types at each...
Fast inference is of paramount value to a wide range deep learning applications. This work presents FTDL, highly-scalable FPGA overlay framework for applications, address the architecture and hardware mismatch faced by traditional efforts. The FTDL specifically optimized tiled structure FPGAs, thereby achieving post-place-and-route operating frequencies exceeding 88 % theoretical maximum across different devices design scales. A flexible compilation efficiently schedules matrix multiply...
By quantizing weights with different precision for parts of a network, mixed-precision quantization promises to reduce the hardware cost and improve speed deep neural network (DNN) accelerators that typically operate fixed scheme. However, additional control needed, decreased efficiency arising from multi-precision operations have made schemes challenging deploy in practice. In this paper, practical framework called MSD leverages heterogeneous computing resources on FPGA perform bit-serial...
Although there are distinct power-performance advantages in customizing an accelerator for a specific combination of FPGA platform and neural network model, developing such highly customized accelerators is challenging task due to the massive design space spans from range models be accelerated, target platform's compute capability, its memory capacity performance characteristics. To address this architectural customization problem, automatic exploration (DSE) framework using mixed-integer...
Low-level warm clouds are a major component in multilayered cloud systems and they generally hidden from the top-down view of satellites with passive measurements. This study conducts an investigation on oceanic embedded structures by using spaceborne radar data fine vertical resolution. The occurrences overlapping geometric features several kinds layers examined. It is found that there three main types involve layers, including single layer clouds, cold-warm double warm-warm clouds. two...
Modern Deep Neural Networks (DNNs) are no-torious for their large memory footprint, which impacts not only the storage capacity requirement in resource-constrained embedded systems, but also performance of an inference machine due to data movement. In this work, we demonstrate a transparent weight compression scheme, called SqueezeBlock, effectively reduces footprint DNN models with minimal impact on accuracy without need retraining. SqueezeBlock employs three steps, namely, clustering,...
Hardware acceleration of deep learning (DL) systems has been increasingly studied to achieve desirable performance and energy efficiency. The FPGA strikes a balance between high efficiency fast development cycle therefore is widely used as DNN accelerator. However, there exists an architecture-layout mismatch in the current designs, which introduces scalability flexibility issues, leading irregular routing resource imbalance problems. To address these limitations, this work, we propose FTDL,...
Satellite rainrate estimation is a great challenge, especially in mesoscale convective systems (MCSs), which mainly due to the absence of direct physical connection between observable cloud parameters and surface rainrate. The machine learning technique was employed this study estimate MCS domain via using top temperature (CTT) derived from geostationary satellite. Five kinds models were investigated, i.e., polynomial regression, support vector machine, decision tree, random forest,...
Event-based vision represents a paradigm shift in how information is captured and processed. By only responding to dynamic intensity changes the scene, event-based sensing produces far less data than conventional frame-based cameras, promising springboard new generation of high-speed, low-power machines for edge intelligence. However, processing such dynamically sparse input originated from event cameras efficiently real time, particularly with complex deep neural networks (DNN), remains...
Fault diagnosis is an effective means to improve the reliability of electric drive system new energy vehicles. At present, inter turn short circuit faults in vehicle motors are mainly detected using single fault features and small sample datasets. This method has low accuracy poor robustness. In this paper, a combined expansion strategy conditional generative adversarial networks for attention mechanism optimization proposed, combining improved Convolutional neural network proposed. First,...
Event-based vision represents a paradigm shift in how information is captured and processed. By only responding to dynamic intensity changes the scene, event-based sensing produces far less data than conventional frame-based cameras, promising springboard new generation of high-speed, low-power machines for edge intelligence. However, processing such dynamically sparse input originated from event cameras efficiently real time, particularly with complex deep neural networks (DNN), remains...
Safe reinforcement learning (RL) is crucial for deploying RL agents in real-world applications, as it aims to maximize long-term rewards while satisfying safety constraints. However, safe often suffers from sample inefficiency, requiring extensive interactions with the environment learn a policy. We propose Efficient Policy Optimization (ESPO), novel approach that enhances efficiency of through manipulation. ESPO employs an optimization framework three modes: maximizing rewards, minimizing...
To accelerate the inference of deep neural networks (DNNs), quantization with low-bitwidth numbers is actively researched. A prominent challenge to quantize DNN models into without significant accuracy degradation, especially at very low bitwidths (< 8 bits). This work targets an adaptive data representation variable-length encoding called DyBit. DyBit can dynamically adjust precision and range separate bit-field be adapted weights/activations distribution. We also propose a hardware-aware...
Fog leads to dangerous driving conditions.Along the highway of coast Bohai gulf, occurrence frequency is higher than other places in Beijing area.Nowadays, numerical weather prediction widely used China.Based on model (MM5), statistical method applied build a forecast for seven stations along highway.When fog it tracked by satellite images.Beijing Weather Information Service provides products each day via various media.These include two forecasts next 24 hr period site highway.Also, there...
&lt;p&gt;Low-level warm clouds are a major component in multilayered cloud systems and generally hidden from the top-down view of satellites with passive measurements. By using spaceborne radar data fine vertical resolution, this study conducts an investigation on oceanic embedded structures. The occurrences overlapping geometric features several kinds layers examined. It is found that there three main types involve layers, including single layer clouds, cold-warm double warm-warm...
Slot filling is a fundamental task in dialog state tracking task-oriented systems. In multi-domain system, user utterances and system responses may mention multiple named entities attributes values. A needs to select those that are confirmed by the fill them into destined slots. One difficulty since dialogue session contains system-user turns, feeding all tokens deep model such as BERT can be challenging due limited capacity of input word GPU memory. this paper, we investigate an...