VALSE Webinar 20230607-12期 总第312期 开放世界的感知：探索可迁移与可持续学习之路 ...

报告嘉宾：刘伊凡 (苏黎世联邦理工)

报告题目：Generalizable, Transferable, and Deployable Perception Systems in the Real World

报告嘉宾：巩东 (澳大利亚新南威尔士大学)

报告题目：Continual Learning and Memory Augmentation with Deep Neural Networks

报告嘉宾：刘伊凡 (苏黎世联邦理工)

报告时间：2023年06月07日 (星期三)晚上20:00 (北京时间)

报告题目：Generalizable, Transferable, and Deployable Perception Systems in the Real World

报告人简介：

Dr. Liu is a researcher at ETH Zurich, a lecturer (a.k.a assistant professor)at the University of Adelaide and a member of the Australian Institution for Machine Learning. Previously, she was a Senior Researcher at the University of Cambridge and received her Ph.D. degree in computer science from the University of Adelaide. She received the Google PhD Fellowship Award for her efficient dense prediction proposal in 2020. She won the `Women in AI’ award of the Defence Artificial Intelligence Research Network and the `Australian Pattern Recognition Society (APRS)ECR Award’ in 2022. Her research interest lies in building robust and generalized vision systems with timely and reliable feedback.

刘伊凡是苏黎世联邦理工大学的研究员，曾是阿德莱德大学计算机科学与工程学院的助理教授和博士生导师，澳大利亚机器学习研究所 (AIML)的研究员。她曾在剑桥大学担任高级研究员，并于2021年12月在阿德莱德大学获得计算机科学博士学位。刘伊凡曾荣获谷歌PhD Fellowship Award (谷歌博士研究奖)，并以计算机视觉为主要研究方向，尤其专注于在真实的开放世界场景中 (例如无人驾驶和机器人等)，提供稳定、可靠、实时的视觉感知，并通过基础模型进行持续学习和快速迁移。

个人主页：

https://irfanicmll.github.io/

报告摘要：

Building an open-world perception system is crucial for real-world applications such as robotics, autonomous driving, and smart agriculture. Unlike closed-world scenarios where AI systems are trained on a fixed set of classes within a given domain, open-world perception requires the model to perform on new domains, new classes and even new tasks that have not been explicitly or fully trained. This talk will delve into our recent work on building open-world perception systems. Our focus will be on creating robust vision systems with large-scale, multi-domain training data, enhancing the ability of trained models to perform new tasks, and transferring knowledge from large models to smaller ones for more timely feedback. Ultimately, our goal is to develop a vision system that can provide precise and timely feedback, enabling a better understanding of objects, scenes, and events in real-world scenarios where the input domain, the number of classes, and the output structures vary a lot and can be constantly changing.

参考文献：

[1] SegViT: Semantic Segmentation with Plain Vision Transformers. Bowen Zhang, Zhi Tian, Quan Tang, Xiangxiang Chu, Xiaolin Wei, Chunhua Shen, Yifan Liu. NeurIPS 2022

[2] ZegCLIP: Towards Adapting CLIP for Zero-shot Semantic Segmentation. Ziqin Zhou, Bowen Zhang, Yinjie Lei, Lingqiao Liu, Yifan Liu. CVPR 2023

报告嘉宾：巩东 (澳大利亚新南威尔士大学)

报告时间：2023年06月07日 (星期三)晚上20:30 (北京时间)

报告题目：Continual Learning and Memory Augmentation with Deep Neural Networks

报告人简介：

Dong Gong is a Lecturer (Assistant Professor)and ARC DECRA Fellow (2023-2026)at the School of CSE,UNSW Sydney. He is also an Adjunct Lecturer with the Australian Institute for Machine Learning (AIML)at The University of Adelaide. Previously, Dong worked as a Research Fellow at the AIML, and a Principal Researcher at the Centre for Augmented Reasoning (CAR), The University of Adelaide. He obtained his Ph.D. degree at Northwestern Polytechnical University in Dec 2018. His research interests are about computer vision and machine learning, especially in learning tasks with non-ideal supervision and flexible requirements in real-world scenarios, such as continual learning.

巩东是新南威尔士大学计算机科学与工程学院的助理教授、博士生导师，澳大利亚机器学习研究所 (AIML)兼职讲师 (阿德莱德大学)。曾在AIML担任研究员，以及阿德莱德大学增强推理中心 (CAR)的首席研究员 (2022年之前)。2018年12月于西北工业大学获得计算机科学博士学位。巩东于2022年获得澳大利亚研究理事会 (Australian Research Council, ARC)优秀青年基金 (Discovery Early Career Researcher Award, DECRA Fellow, 2023-2026)。主要研究计算机视觉和机器学习，特别是在现实场景中具有非理想监督和灵活要求的学习任务，如持续学习等。

个人主页：

https://donggong1.github.io/

报告摘要：

Deep learning (DL)has been successful in many applications. However, the conventional DL approaches focus on the end results on fixed datasets and fail to handle the dynamically raised novel requirements in the real world. Continual learning aims to train deep neural networks (DNNs) to efficiently accumulate knowledge on dynamically arriving data and task streams like humans. The main challenges include how to enable DNNs to learn on various data distributions and tasks without catastrophic forgetting, and how to help a learning system identify novel concepts. To let DNNs properly handle the knowledge learned in the data/ task streams, we think about the problem in view of augmenting the memorization ability of DNNs. We will also discuss how we may use pre-trained models as foundational memory and adapt the learned knowledge in a continual learning process.

参考文献：

[1] Learning Bayesian Sparse Networks with Full Experience Replay for Continual Learning, Qingsen Yan*, Dong Gong*#, Yuhang Liu, Anton van den Hengel, Javen Qinfeng Shi. CVPR, 2022.

[2] Memorizing Normality to Detect Anomaly: Memory-augmented Deep Autoencoder for Unsupervised Anomaly Detection. Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, Anton van den Hengel. ICCV, 2019.

主持人：陆昊 (华中科技大学)

主持人简介：

陆昊，博士，华中科技大学副教授，硕士生导师，武汉英才，楚天学子，中国地质大学 (武汉)本科 (2013)，华中科技大学博士 (2018)，澳大利亚阿德莱德大学访问学生 (2016-2017)与博士后 (2018-2020)，合作导师沈春华教授，2020年11月起任华中科技大学人工智能与自动化学院副教授。回国以来，先后主持或参与了国家自然科学基金青年项目、国家自然科学基金面上项目、科技部重点研发计划等多个课题。陆昊副教授一直从事计算机视觉稠密预测方向的研究工作，在通用稠密预测模型组件设计与优化，特定稠密预测任务 (如图像抠图、目标计数、图像分割、特征匹配等)中的表征与模型设计具有一定研究基础。近年来以一作或者通讯作者在人工智能相关顶级国际会议如NeurIPS、CVPR、ICCV、ECCV、AAAI、ACM MM和人工智能权威期刊如IEEE TPAMI、IJCV、IEEE TIP、IEEE TNNLS、IEEE TCSVT、PR等发表论文20余篇，2篇一作论文入选ESI高被引论文。

个人主页：

https://sites.google.com/site/poppinace/home

特别鸣谢本次Webinar主要组织者：

主办AC：陆昊 (华中科技大学)

协办AC：胡鹏 (四川大学)