A5079738111- Advanced Neural Network Applications
- Domain Adaptation and Few-Shot Learning
- Multimodal Machine Learning Applications
- Advanced Image and Video Retrieval Techniques
- Robotics and Sensor-Based Localization
- Marine Toxins and Detection Methods
- Adversarial Robustness in Machine Learning
- Advanced Wireless Communication Techniques
- Generative Adversarial Networks and Image Synthesis
- Digital Media Forensic Detection
- Wireless Communication Networks Research
- Advanced biosensing and bioanalysis techniques
- Oxidative Organic Chemistry Reactions
- Advanced MIMO Systems Optimization
- Anomaly Detection Techniques and Applications
- Remote-Sensing Image Classification
- Advanced Biosensing Techniques and Applications
- Synthesis of Indole Derivatives
- Synthetic Organic Chemistry Methods
- Biosensors and Analytical Detection
- COVID-19 diagnosis using AI
- Machine Learning and Data Classification
- Immunotherapy and Immune Responses
- Advanced Vision and Imaging
- Microfluidic and Bio-sensing Technologies
École Polytechnique Fédérale de Lausanne
2020-2025
Georgia Institute of Technology
2018-2024
Atlanta Technical College
2022
Nanhua University
2022
Meta (Israel)
2021
National Dong Hwa University
2020
National Tsing Hua University
2011-2019
National Cheng Kung University
2009-2019
National Taiwan University
2011-2018
National Yang Ming Chiao Tung University
2009-2013
Few-shot classification aims to learn a classifier recognize unseen classes during training with limited labeled examples. While significant progress has been made, the growing complexity of network designs, meta-learning algorithms, and differences in implementation details make fair comparison difficult. In this paper, we present 1) consistent comparative analysis several representative few-shot results showing that deeper backbones significantly reduce performance among methods on...
Monocular depth estimation is a challenging task in scene understanding, with the goal to acquire geometric properties of 3D space from 2D images. Due lack RGB-depth image pairs, unsupervised learning methods aim at deriving information alternative supervision such as stereo pairs. However, most existing works fail model structure objects, which generally results considering pixel-level objective functions during training. In this paper, we propose SceneNet overcome limitation aid semantic...
Fully functional, tumor-specific, SLAMF7-expressing cytotoxic CD4 T cells directly mediate immunity against human cancer.
Continual learning has received a great deal of attention recently with several approaches being proposed. However, evaluations involve diverse set scenarios making meaningful comparison difficult. This work provides systematic categorization the and evaluates them within consistent framework including strong baselines state-of-the-art methods. The results provide an understanding relative difficulty that simple (Adagrad, L2 regularization, naive rehearsal strategies) can surprisingly...
Semi-supervised learning, i.e., training networks with both labeled and unlabeled data, has made significant progress recently. However, existing works have primarily focused on image classification tasks neglected object detection which requires more annotation effort. In this work, we revisit the Semi-Supervised Object Detection (SS-OD) identify pseudo-labeling bias issue in SS-OD. To address this, introduce Unbiased Teacher, a simple yet effective approach that jointly trains student...
We address the task of domain adaptation in object detection, where there is an obvious gap between a with annotations (source) and interest without (target). As popular semi-supervised learning method, teacher-student framework (a student model supervised by pseudo labels from teacher model) has also yielded large accuracy gain cross-domain detection. However, it suffers shift generates many low-quality (e.g., false positives), which leads to sub-optimal performance. To mitigate this...
With the recent development of Semi-Supervised Object Detection (SS-OD) techniques, object detectors can be improved by using a limited amount labeled data and abundant unlabeled data. However, there are still two challenges that not addressed: (1) is no prior SS-OD work on anchor-free detectors, (2) works ineffective when pseudo-labeling bounding box regression. In this paper, we present Unbiased Teacher v2, which shows generalization method to also introduces Listen2Student mechanism for...
Person re-identification (Re-ID) aims at recognizing the same person from images taken across different cameras. To address this task, one typically requires a large amount labeled data for training an effective Re-ID model, which might not be practical real-world applications. alleviate limitation, we choose to exploit sufficient of pre-existing (auxiliary) dataset. By jointly considering such auxiliary dataset and interest (but without label information), our proposed adaptation network...
While significant advances have been made for single-agent perception, many applications require multiple sensing agents and cross-agent communication due to benefits such as coverage robustness. It is therefore critical develop frameworks which support multi-agent collaborative perception in a distributed bandwidth-efficient manner. In this paper, we address the problem, where one agent required perform task can communicate share information with other on same task. Specifically, propose...
While representation learning aims to derive interpretable features for describing visual data, disentanglement further results in such so that particular image attributes can be identified and manipulated. However, one cannot easily address this task without observing ground truth annotation the training data. To problem, we propose a novel deep model of Cross-Domain Representation Disentangler (CDRD). By fully annotated source-domain data unlabeled target-domain interest, our bridges...
We present a novel and unified deep learning framework which is capable of domain-invariant representation from data across multiple domains. Realized by adversarial training with additional ability to exploit domain-specific information, the proposed network able perform continuous cross-domain image translation manipulation, produces desirable output images accordingly. In addition, resulting feature exhibits superior performance unsupervised domain adaptation, also verifies effectiveness...
In this paper, we propose the problem of collaborative perception, where robots can combine their local observations with those neighboring agents in a learnable way to improve accuracy on perception task. Unlike existing work robotics and multi-agent reinforcement learning, formulate as one learned information must be shared across set bandwidth-sensitive manner optimize for scene understanding tasks such semantic segmentation. Inspired by networking communication protocols, multi-stage...
Methods for the analysis of cell secretions at single-cell level only provide semiquantitative endpoint readouts. Here we describe a microwell array real-time spatiotemporal monitoring extracellular from hundreds single cells in parallel. The incorporates gold substrate with arrays nanometric holes functionalized receptors specific analyte, and is illuminated light spectrally overlapping device's spectrum extraordinary optical transmission. Spectral shifts surface plasmon resonance resulting...
Using multiple spatial modalities has been proven helpful in improving semantic segmentation performance. However, there are several real-world challenges that have yet to be addressed: (a) label efficiency and (b) enhancing robustness realistic scenarios where missing at the test time. To address these challenges, we first propose a simple efficient multi-modal fusion mechanism Linear Fusion, performs better than state-of-the-art models even with limited supervision. Second, M3L:...
Organoid tumor models have emerged as a powerful tool in the fields of biology and medicine such 3D structures grown from cells recapitulate better characteristics, making these tumoroids unique for personalized cancer research. Assessment their functional behavior, particularly protein secretion, is significant importance to provide comprehensive insights. Here, label-free spectroscopic imaging platform presented with advanced integrated optofluidic nanoplasmonic biosensor that enables...
Land cover classification aims at classifying each pixel in a satellite image into particular land category, which can be regarded as multi-class semantic segmentation task. In this paper, we propose deep aggregation network for solving task, extracts and combines multi-layer features during the process. particular, introduce soft labels graph-based fine tuning our proposed improving performance. experiments, demonstrate that performs favorably against state-of-the-art models on dataset of...
Current imaging technologies are limited in their capability to simultaneously capture intracellular and extracellular dynamics a spatially temporally resolved manner. This study presents multimodal system that integrates nanoplasmonic sensing with multichannel fluorescence concomitantly analyze processes space time at the single-cell level. Utilizing highly sensitive gold nanohole array biosensor, provides label-free real-time monitoring of secretion, while implementing...
Although known for over a quarter of century, the oxidative radical cyclisation route to spiroketals has found limited use in natural product synthesis comparison classical approaches. Its successful application this field research forms subject perspective.
Recent studies on transfer learning have shown that selectively fine-tuning a subset of layers or customizing different rates for each layer can greatly improve robustness to out-of-distribution (OOD) data and retain generalization capability in the pre-trained models. However, most these methods employ manually crafted heuristics expensive hyper-parameter searches, which prevent them from scaling up large datasets neural networks. To solve this problem, we propose Trainable Projected...
Neural Networks can perform poorly when the training label distribution is heavily imbalanced, as well testing data differs from distribution. In order to deal with shift in distribution, which imbalance causes, we motivate problem perspective of an optimal Bayes classifier and derive a post-training prior rebalancing technique that be solved through KL-divergence based optimization. This method allows flexible hyper-parameter efficiently tuned on validation set effectively modify margin...
Adapting large-scale pretrained models to various downstream tasks via fine-tuning is a standard method in machine learning. Recently, parameter-efficient methods show promise adapting model different while training only few parameters. Despite their success, most existing are proposed Natural Language Processing with language Transformers, and adaptation Computer Vision Transformers remains under-explored, especially for dense vision tasks. Further, multi-task settings, individually storing...
In this paper, we propose three classes of systematic approaches for constructing zero correlation zone (ZCZ) sequence families. most cases, these are capable generating families that achieve the upper bounds on family size ($K$) and ZCZ width ($T$) a given period ($N$). Our can produce various binary polyphase with desired parameters $(N,K,T)$ alphabet size. They also provide additional tradeoffs amongst above four system less constrained by Furthermore, constructed have nested-like...