A5055016576- Advanced Neural Network Applications
- Domain Adaptation and Few-Shot Learning
- Multimodal Machine Learning Applications
- Advanced Image and Video Retrieval Techniques
- Adversarial Robustness in Machine Learning
- COVID-19 diagnosis using AI
- Underwater Vehicles and Communication Systems
- Indoor and Outdoor Localization Technologies
- Energy Efficient Wireless Sensor Networks
- Machine Learning and Data Classification
- Image Processing Techniques and Applications
- Brain Tumor Detection and Classification
- Recommender Systems and Techniques
- Energy Harvesting in Wireless Networks
- Image Enhancement Techniques
- Network Time Synchronization Technologies
- Topic Modeling
- Advanced Graph Neural Networks
- Anomaly Detection Techniques and Applications
- Generative Adversarial Networks and Image Synthesis
- Neural Networks and Applications
- Underwater Acoustics Research
- Machine Learning and ELM
- Nonlinear Dynamics and Pattern Formation
- Mobile Ad Hoc Networks
Chinese Academy of Sciences
2015-2025
Institute of Computing Technology
2015-2025
University of Chinese Academy of Sciences
2010-2023
Beijing Tian Tan Hospital
2023
Capital Medical University
2023
Xiamen University
2023
Hefei University of Technology
2004-2006
Current Knowledge Distillation (KD) methods for semantic segmentation often guide the student to mimic teacher's structured information generated from individual data samples. However, they ignore global relations among pixels across various images that are valuable KD. This paper proposes a novel Cross-Image Relational KD (CIRKD), which focuses on transferring pixel-to-pixel and pixel-to-region whole images. The motivation is good teacher network could construct well-structured feature...
Pathogenic bioaerosols are critical for outbreaks of airborne disease; however, rapidly and accurately identifying pathogens directly from complex air environments remains highly challenging. We present an advanced method that combines open-set deep learning (OSDL) with single-cell Raman spectroscopy to identify in real-world containing diverse unknown indigenous bacteria cannot be fully included training sets. To test further enhance identification, we constructed the datasets aerosolized...
We present a collaborative learning method called Mutual Contrastive Learning (MCL) for general visual representation learning. The core idea of MCL is to perform mutual interaction and transfer contrastive distributions among cohort networks. A crucial component Interactive (ICL). Compared with vanilla learning, ICL can aggregate cross-network embedding information maximize the lower bound between two This enables each network learn extra knowledge from others, leading better feature...
The teacher-free online Knowledge Distillation (KD) aims to train an ensemble of multiple student models collaboratively and distill knowledge from each other. Although existing KD methods achieve desirable performance, they often focus on class probabilities as the core type, ignoring valuable feature representational information. We present a Mutual Contrastive Learning (MCL) framework for KD. idea MCL is perform mutual interaction transfer contrastive distributions among cohort networks...
Knowledge distillation often involves how to define and transfer knowledge from teacher student effectively. Although recent self-supervised contrastive achieves the best performance, forcing network learn such may damage representation learning of original class recognition task. We therefore adopt an alternative augmented task guide joint distribution auxiliary It is demonstrated as a richer improve power without losing normal classification capability. Moreover, it incomplete that...
Class incremental learning (CIL) aims to solve the notorious forgetting problem, which refers fact that once network is updated on a new task, its performance previously-learned tasks degenerates catastrophically. Most successful CIL methods store exemplars (samples of learned tasks) train feature extractor incrementally, or prototypes (features estimate distribution. However, stored would violate data privacy concerns, while fixed might not reasonably be consistent with distribution,...
Underwater Sensor Network (UWSN) is a representative three-dimensional wireless sensor network. Due to the unique characteristics of underwater acoustic communication, providing energy-efficient and low-latency routing protocols for UWSNs challenging. Major challenges are water currents, limited resources, long propagation delay. topology dynamic complex as sensors have always been moving with currents. Some proposed adopt geographic address this problem, but localization hard obtain in...
Latest algorithms for automatic neural architecture search perform remarkable but are basically directionless in space and computational expensive the training of every intermediate architecture. In this paper, we propose a method efficient called EENA (Efficient Evolution Neural Architecture). Due to elaborately designed mutation crossover operations, evolution process can be guided by information have already been learned. Therefore, less effort will required while searching time reduced...
We propose a simple yet effective method to reduce the redundancy of DenseNet by substantially decreasing number stacked modules replacing original bottleneck our SMG module, which is augmented local residual. Furthermore, module equipped with an efficient two-stage pipeline, aims DenseNet-like architectures that need integrate all previous outputs, i.e., squeezing incoming informative but redundant features gradually hierarchical convolutions as hourglass shape and then exciting it...
We proposed a Prior Gradient Mask Guided Pruning-aware Fine-Tuning (PGMPF) framework to accelerate deep Convolutional Neural Networks (CNNs). In detail, the PGMPF selectively suppresses gradient of those ”unimportant” parameters via prior mask generated by pruning criterion during fine-tuning. has three charming characteristics over previous works: (1) network A typical pipeline consists training, and fine-tuning, which are relatively independent, while utilizes variant as guide without...
Unlike radio communication in terrestrial wireless sensor networks, acoustic channel underwater network (UWSN) is challenged by long propagation delay and swarm mobility. The causes the deterioration of throughput unfairness. Moreover, mobility unreliable access failure reservation. In this paper, we propose a novel tolerant MAC protocol (DTMAC) inspired coupon collector's problem. It proposes distributed collection algorithm UWSNs. If node needs to send packet, packet will be repeatedly...
Due to long propagation delays, routing discovery is very expensive in underwater acoustic sensor networks. Therefore, depth-based (DBR) protocols are usually preferred whose cost almost zero. However, DBR working with broadcast medium access control (MAC) bring severe collisions especially the data collection network. Moreover, DBR, all packets forwarded from deep shallow nodes along direction toward sink node. The directional packet transmissions may lead load imbalance, where some...
Multi-teacher Knowledge Distillation (KD) transfers diverse knowledge from a teacher pool to student network. The core problem of multi-teacher KD is how balance distillation strengths among various teachers. Most existing methods often develop weighting strategies an individual perspective performance or teacher-student gaps, lacking comprehensive information for guidance. This paper proposes Multi-Teacher with Reinforcement Learning (MTKD-RL) optimize weights. In this framework, we...
In safety-critical domains such as medical diagnostics and autonomous driving, single-image evidence is sometimes insufficient to reflect the inherent ambiguity of vision problems. Therefore, multiple plausible assumptions that match image semantics may be needed actual distribution targets support downstream tasks. However, balancing improving diversity consistency segmentation predictions under high-dimensional output spaces potential multimodal distributions still challenging. This paper...
Diffusion models have received wide attention in generation tasks. However, the expensive computation cost prevents application of diffusion resource-constrained scenarios. Quantization emerges as a practical solution that significantly saves storage and by reducing bit-width parameters. existing quantization methods for still cause severe degradation performance, especially under extremely low bit-widths (2-4 bit). The primary decrease performance comes from significant discretization...
Similar to other cyber infrastructure systems, as wireless sensor networks become larger and more complex, many classic algorithms may no longer work efficiently. This paper presents a network time synchronization model that was initially inspired by synchronous flashing of fireflies. Synchronous fireflies is an interesting phenomenon has been studied for decades. A variety models have proposed explain this phenomenon, among which the pulse-coupled oscillators oscillators. The in such...