A5056949827- Advanced Neural Network Applications
- Parallel Computing and Optimization Techniques
- Advanced Memory and Neural Computing
- Adversarial Robustness in Machine Learning
- Ferroelectric and Negative Capacitance Devices
- Reinforcement Learning in Robotics
- Domain Adaptation and Few-Shot Learning
- Embedded Systems Design Techniques
- Neural Networks and Applications
- CCD and CMOS Imaging Sensors
- Machine Learning and Data Classification
- Topic Modeling
- Anomaly Detection Techniques and Applications
- Evolutionary Algorithms and Applications
- Radiation Effects in Electronics
- Natural Language Processing Techniques
- Multimodal Machine Learning Applications
- Advanced Image and Video Retrieval Techniques
- Advanced Graph Neural Networks
- Graph Theory and Algorithms
- Model Reduction and Neural Networks
- VLSI and Analog Circuit Testing
- Generative Adversarial Networks and Image Synthesis
- Interconnection Networks and Systems
- Human Pose and Action Recognition
Chinese Academy of Sciences
2015-2025
Institute of Computing Technology
2015-2025
Shanghai Innovative Research Center of Traditional Chinese Medicine
2024
Tsinghua University
2024
Cambricon (China)
2016-2023
State Key Laboratory of Computer Architecture
2015-2016
University of Chinese Academy of Sciences
2015-2016
International Centre for Theoretical Physics Asia-Pacific
2014
Machine-Learning tasks are becoming pervasive in a broad range of domains, and systems (from embedded to data centers). At the same time, small set machine-learning algorithms (especially Convolutional Deep Neural Networks, i.e., CNNs DNNs) proving be state-of-the-art across many applications. As architectures evolve towards heterogeneous multi-cores composed mix cores accelerators, accelerator can achieve rare combination efficiency (due number target algorithms) application scope.
In recent years, neural network accelerators have been shown to achieve both high energy efficiency and performance for a broad application scope within the important category of recognition mining applications.
Neural networks (NNs) have been demonstrated to be useful in a broad range of applications such as image recognition, automatic translation and advertisement recommendation. State-of-the-art NNs are known both computationally memory intensive, due the ever-increasing deep structure, i.e., multiple layers with massive neurons connections (i.e., synapses). Sparse neural emerged an effective solution reduce amount computation required. Though existing NN accelerators able efficiently process...
Neural networks (NNs) have been demonstrated to be useful in a broad range of applications such as image recognition, automatic translation and advertisement recommendation. State-of-the-art NNs are known both computationally memory intensive, due the ever-increasing deep structure, i.e., multiple layers with massive neurons connections (i.e., synapses). Sparse neural emerged an effective solution reduce amount computation required. Though existing NN accelerators able efficiently process...
Machine-Learning tasks are becoming pervasive in a broad range of domains, and systems (from embedded to data centers). At the same time, small set machine-learning algorithms (especially Convolutional Deep Neural Networks, i.e., CNNs DNNs) proving be state-of-the-art across many applications. As architectures evolve towards heterogeneous multi-cores composed mix cores accelerators, accelerator can achieve rare combination efficiency (due number target algorithms) application scope. Until...
Machine-Learning tasks are becoming pervasive in a broad range of domains, and systems (from embedded to data centers). At the same time, small set machine-learning algorithms (especially Convolutional Deep Neural Networks, i.e., CNNs DNNs) proving be state-of-the-art across many applications. As architectures evolve towards heterogeneous multi-cores composed mix cores accelerators, accelerator can achieve rare combination efficiency (due number target algorithms) application scope. Until...
Neural Networks (NN) are a family of models for broad range emerging machine learning and pattern recondition applications. NN techniques conventionally executed on general-purpose processors (such as CPU GPGPU), which usually not energy-efficient since they invest excessive hardware resources to flexibly support various workloads. Consequently, application-specific accelerators neural networks have been proposed recently improve the energy-efficiency. However, such were designed small set...
Neural networks have become the dominant algorithms rapidly as they achieve state-of-the-art performance in a broad range of applications such image recognition, speech recognition and natural language processing. However, neural keep moving towards deeper larger architectures, posing great challenge to huge amount data computations. Although sparsity has emerged an effective solution for reducing intensity computation memory accesses directly, irregularity caused by (including sparse...
In recent years, neural network accelerators have been shown to achieve both high energy efficiency and performance for a broad application scope within the important category of recognition mining applications. Still, such remain limited by memory accesses. this paper, we focus on image applications, arguably most among The networks which are state-of-the-art these applications Convolutional Neural Networks (CNN), they an property: weights shared many neurons, considerably reducing...
In recent years, inexact computing has been increasingly regarded as one of the most promising approaches for reducing energy consumption in many applications that can tolerate a degree inaccuracy. Driven by principle trading tolerable amounts application accuracy return significant resource savings - consumed, (critical path) delay and (silicon) area being resources this approach limited to certain domains. paper, we propose expand scope, error tolerance well systems through neural network...
Neural Networks (NN) are a family of models for broad range emerging machine learning and pattern recondition applications. NN techniques conventionally executed on general-purpose processors (such as CPU GPGPU), which usually not energy-efficient since they invest excessive hardware resources to flexibly support various workloads. Consequently, application-specific accelerators neural networks have been proposed recently improve the energy-efficiency. However, such were designed small set...
A vast array of devices, ranging from industrial robots to self-driven cars or smartphones, require increasingly sophisticated processing real-world input data (image, voice, radio, ...). Interestingly, hardware neural network accelerators are emerging again as attractive candidate architectures for such tasks. The algorithms considered come two, largely separate, domains: machine-learning and neuroscience. These networks have very different characteristics, so it is unclear which approach...
Continuous-valued deep convolutional networks (DNNs) can be converted into accurate rate-coding based spike neural (SNNs). However, the substantial computational and energy costs, which is caused by multiple spikes, limit their use in mobile embedded applications. And recent works have shown that newly emerged temporal-coding SNNs from DNNs reduce load effectively. In this paper, we propose a novel method to convert SNNs, called TDSNN. Combined with characteristic of leaky integrate-andfire...
In the field of natural language processing, rapid development large model (LLM) has attracted more and attention. LLMs have shown a high level creativity in various tasks, but methods for assessing such are inadequate. The assessment LLM needs to consider differences from humans, requiring multi-dimensional measurement while balancing accuracy efficiency. This paper aims establish an efficient framework LLMs. By adapting modified Torrance Tests Creative Thinking, research evaluates creative...
The graph convolutional network (GCN) emerges as a promising direction to learn the inductive representation in data commonly used widespread applications, such E-commerce, social networks, and knowledge graphs. However, learning from graphs is nontrivial because of its mixed computation model involving both analytics neural computing. To this end, we decompose GCN into two hierarchical paradigms: 1) graph-level 2) node-level Such paradigm facilitates software hardware accelerations for...
In recent years, inexact computing has been increasingly regarded as one of the most promising approaches for slashing energy consumption in many applications that can tolerate a certain degree inaccuracy. Driven by principle trading tolerable amounts application accuracy return significant resource savings-the consumed, (critical path) delay, and (silicon) area-this approach limited to application-specified integrated circuits (ASICs) so far. These ASIC realizations have narrow scope are...
Recent emerged quantization technique (i.e., using low bit-width fixed-point data instead of high floating-point data) has been applied to inference deep neural networks for fast and efficient execution. However, directly applying in training can cause significant accuracy loss, thus remaining an open challenge. In this paper, we propose a novel approach, which applies layer-wise precision-adaptive networks. The new approach leverages our key insight that the degradation is attributed...
Deep Learning Accelerators (DLAs) are effective to improve both performance and energy efficiency of compute-intensive deep learning algorithms. A flexible portable mean exploit DLAs is using high-performance software libraries with well-established APIs, which typically either manually implemented or automatically generated by exploration-based compilation approaches. Though approaches significantly reduce programming efforts, they fail find optimal near-optimal programs from a large but...
Recent advancements in open-source code large language models (LLMs) have been driven by fine-tuning on the data generated from powerful closed-source LLMs, which are expensive to obtain. This paper explores whether it is possible use a fine-tuned model generate additional augment its instruction-tuning dataset. We make two observations: (1) A snippet can serve as response different instructions. (2) Instruction-tuned LLMs perform better at translating into instructions than reverse. Based...