Section: Research Program

High dimensional data processing and search

High dimensionality is inherent in applications involving images, audio and text as well as in many scientific applications involving raster data or high-throughput data. Because of the dimensionality curse, technologies for processing and analyzing such data cannot rely on traditional relational DBMS or data mining methods. It rather requires to employ machine learning methods such as dimensionality reduction, representation learning or random projection. The activity of Zenith in this domain focuses on methods that permit data processing and search at scale, in particular in the presence of strong uncertainty and/or ambiguity. Actually, while small datasets are often characterized by a careful collection process, massive amounts of data often come with outliers and spurrious items, because it appears impossible to guarantee faultless collection at massive bandwidth. Another source of noise is often the sensor itself, that may be of low quality but of high sampling rate, or even the actual content, e.g. in cultural heritage applications when historical content appears seriously damaged by time. To attack these difficult problems, we focus on the following research topics:

• Uncertainty estimation. Items in massive datasets may either be uncertain, e.g. for automatically annotated data as in image analysis, or be more or less severely corrupted by noise, e.g. in noisy audio recordings or in the presence of faulty sensors. In both cases, the concept of uncertainty is central for the end-user to exploit the content. In this context, we use probability theory to quantify uncertainty and propose machine learning algorithms that may operate robustly, or at least assess the quality of their output. This vast topic of research is guided by large-scale applications (both data search and data denoising), and our research is oriented towards computationally effective methods.

• Deep neural networks. A major breakthrough in machine learning performance has been the advent of deep neural networks, which are characterized by high numbers (millions) of parameters and scalable learning procedures. We are striving towards original architectures and methods that are theoretically grounded and offer state-of-the-art performance for data search and processing. The specific challenges we investigate are: very high dimensionality for static data and very long-term dependency for temporal data, both in the case of possibly strong uncertainty or ambiguity (e.g. hundreds of thousands of classes).

• Community service. Research in machine learning is guided by applications. In Zenith, two main communities are targeted: botany, and digital humanities. In both cases, our observation is that significant breakthroughs may be achieved by connecting these communities to machine learning researchers. This may be achieved through wording application-specific problems in classical machine learning parlance. Thus, the team is actively involved in the organization of international evaluation campaigns that allow machine learning researchers to propose new methods while solving important application problems.