5月9日学术报告： Guillaume Gravier & Laurent Amsaleg

报告人：Guillaume Gravier、 Laurent Amsaleg

主持人：王长波

报告时间：5月9日（周二）10:00-11:30

报告地点：中北校区数学馆201

Part 1

报告题目：Structuring multimedia collections with links for media analytic

报告人：Guillaume Gravier

报告摘要：

Media analytics refers to interactive exploration of multimedia collections to search for information and gain insight on a topic of interest. Exploring large-scale collections requires an organization of the collection in addition to the description of each item, the latter being mostly relevant for search but not for exploration. We investigate graph-based models of collection, focusing on the hyperlinking step in videos and in news collections. In this talk, I will review recent research activities on the topic in the Linkmedia team at Irisa, Rennes, France. We will discuss several techniques for multimedia hyperlinking, e.g., hierarchical topic models, multimodal topic models, multimodal embeding with symmetrical neural networks. We will also discuss user acceptability of hyperlink navigation, in particular describing a novel nearest neighbor graph construction algorithm with good navigability properties. Recent user tests demonstrated the interest of this type of graphs for news analytics. We will conclude with perspective on a use-case in data journalism that require hyperlinking heterogeneous information sources beyond multimedia content.

报告人简介：

Guillaume Gravier is a CNRS senior research scientist at Irisa where he heads the Linkmedia research team (http://www-linkmedia.irisa.fr). His research interests are in multimedia analytics, ranging from content description and retrieval to information extraction and navigation interfaces. He has a background in audio, speech and language processing in multimedia and has been involved in the development of many multimedia systems within the framework of international benchmarks (speech transcription, video geolocation, video hyperlinking, spoken web search, speaker naming, etc.).

Part 2

报告题目：Image recognition and adversarial attacks

报告人：Laurent Amsaleg

报告摘要：

Recent research has shown that image recognition is vulnerable to adversarial attacks. By adding to the query image an imperceptible amount of adversarial noise, it is highly likely that image recognition can be tricked, possibly returning with a high similarity confidence an image that we, as humans, find extremely dissimilar. Such adversarial behaviors have also been observed in the context of machine learning and deep learning. We will discuss these issues.

报告人简介：

Laurent Amsaleg is a research scientist at CNRS at Irisa in the Linkmedia research team. He worked on relational and object-oriented databases, garbage collection, micro-kernels and single-level stores as well as flexible database query execution strategies. His research now focuses on very large scale high-dimensional indexing as well as on the security and privacy dimensions of multimedia contents.