A5074071197- Advanced Neural Network Applications
- Domain Adaptation and Few-Shot Learning
- Neural Networks and Applications
- Advanced Vision and Imaging
- Parallel Computing and Optimization Techniques
- Advanced Graph Neural Networks
- 3D Shape Modeling and Analysis
- VLSI and Analog Circuit Testing
- Computer Graphics and Visualization Techniques
- Advancements in Photolithography Techniques
- Advanced Image and Video Retrieval Techniques
- Integrated Circuits and Semiconductor Failure Analysis
- Multimodal Machine Learning Applications
- Advanced Memory and Neural Computing
- Machine Learning and ELM
- Image and Object Detection Techniques
- Quantum-Dot Cellular Automata
- Graph Theory and Algorithms
- Generative Adversarial Networks and Image Synthesis
- Computational Physics and Python Applications
- Advanced Image Processing Techniques
- VLSI and FPGA Design Techniques
- Image and Video Quality Assessment
- Interconnection Networks and Systems
- CCD and CMOS Imaging Sensors
Institute of Computing Technology
2021-2025
Chinese Academy of Sciences
2021-2025
Cambricon (China)
2021
University of Chinese Academy of Sciences
2021
Tsinghua University
2017
Graph neural networks (GNNs), which extend traditional for processing graph-structured data, have been widely used in many fields. The GNN computation mainly consists of the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">edge processing</i> to generate messages by combining edge/vertex features and xmlns:xlink="http://www.w3.org/1999/xlink">vertex update vertex with aggregated messages. In addition nontrivial vector operations edge...
Deep neural network (DNN) training is notoriously time-consuming, and quantization promising to improve the efficiency with reduced bandwidth/storage requirements computation costs. However, state-of-the-art quantized algorithms negligible accuracy loss, which require on-the-fly statistic-based over a great amount of data (e.g., neurons weights) high-precision weight update, cannot be effectively deployed on existing DNN accelerators. To address this problem, we propose first customized...
Domain adaptive object detection (DAOD) aims to infer a robust detector on the target domain with labelled source datasets. Recent studies utilize feature extractor shared and domains capture domain-invariant features task-relevant information both feature-alignment constraint annotations. However, across discards partial of due gap lack annotations, leading compromised discrimination capabilities within domain. To this end, we propose novel REmainder Adaptive CompensaTion network (REACT)...
Domain adaptive object detection (DAOD) aims to generalize detectors trained on an annotated source domain unlabelled target domain. However, existing methods focus reducing the bias of backbone by inferring a discriminative visual encoder, while ignoring in head. Inspired high generalization vision-language models (VLMs), applying VLM as robust following domain-aware head is reasonable way learn detector for each domain, rather than traditional methods. To achieve above issue, we thus...
Domain adaptive object detection (DAOD) aims to generalize detectors trained on an annotated source domain unlabelled target domain. As the visual-language models (VLMs) can provide essential general knowledge unseen images, freezing visual encoder and inserting a domain-agnostic adapter learn domain-invariant for DAOD. However, is inevitably biased It discards some beneficial discriminative domain, i.e., domain-specific of To solve issue, we propose novel Domain-Aware Adapter (DA-Ada)...
Unary computing, whose arithmetics require only one logic gate, has enabled efficient DNN processing, especially on strictly power-constrained devices. However, unary computing still confronts the power efficiency bottleneck for buffering bitstreams. The of bitstreams requires accumulating bits into large bitwidth binary numbers. number needs to activate all per cycle in case carry propagation. As a result, accumulation process accounts 32%-70% budget.
Neural scene representation (NSR) initiates a new methodology of encoding 3D with neural networks by learning from dozens photos taken different camera positions. NSR not only achieves significant improvement in the quality novel view synthesis and reconstruction but also reduces cost expensive laser cameras to cheap color on shelf. However, performing using is far real-time due extremely low hardware utilization (only peak performance), which greatly limits its applications AR/VR interactions
Research on emergent communication has recently gained significant traction as a promising avenue for the linguistic community to unravel human language's origins and explore artificial intelligence's generalization capabilities. Current research predominantly concentrated recognizing qualitative patterns of object attributes(e.g., shape color) paid little attention quantitative relationship among quantities which is known part numerical concepts. The ability generalize concepts, i.e.,...
Heterogeneous collaborative computing with NPU and CPU has received widespread attention due to its substantial performance benefits.To ensure data confidentiality integrity during computing, Trusted Execution Environments (TEE) is considered a promising solution because of comparatively lower overhead.However, existing heterogeneous TEE designs are inefficient for fine different memory granularities between NPU. 1) The cacheline granularity intensifies pressure extra access, 2) the MAC...
Deploying advanced large language models on edge devices, such as smartphones and robotics, is a growing trend that enhances user data privacy network connectivity resilience while preserving intelligent capabilities. However, task exhibits single-batch computing with incredibly low arithmetic intensity, which poses the significant challenges of huge memory footprint bandwidth demands limited resources. To address these issues, we introduce Cambricon-LLM, chiplet-based hybrid architecture...
In the current landscape, high-resolution (HR) videos have gained immense popularity, promising an elevated viewing experience. Recent research has demonstrated that video super-resolution (SR) algorithm, empowered by deep neural networks (DNNs), can substantially enhance quality of HR processing low-resolution (LR) frames. However, existing DNN models demand significant computational resources, posing challenges for deployment SR algorithms on client devices. While numerous accelerators...
Research on emergent communication between deep-learning-based agents has received extensive attention due to its inspiration for linguistics and artificial intelligence. However, previous attempts have hovered around emerging under perception-oriented environmental settings, that forces describe low-level perceptual features intra image or symbol contexts. In this work, inspired by the classic human reasoning test (namely Raven's Progressive Matrix), we propose Reasoning Game, a...