Esteban Marquer (Orpailleur) will defend his thesis, entitled “Reasoning over Data: Analogy-based and Transfer Learning to improve Machine Learning”, on Monday, June 24 at 2 p.m., in room B013.
Recent years have seen a renewed interest in the potential of analogy detection and analogical inference, with successful applications in Machine Learning (ML) to the retrieval and generation of images, of text, and structured data such as knowledge graphs, but also the detection of relations between and within images, texts, and structured data. While some of those works are based on an intuitive understanding of analogy, significant effort has been made since the antiquity to define analogies as accurately as possible. Analogies are a key component of human cognition, and can be viewed as an abstraction mechanism that identifies similarities and differences between different situations, and as a reasoning tool to adapt known solutions to new situation. When reasoning by analogy, the goal is to adapt the solution of a known or source problem, which is sufficiently similar to the actual or target problem. This process involves a transfer between the context of the source problem (the problem and its solution) and the context of the target problem. In the past 50 years, different aspects of the notion of Analogical Proportions (APs) have been explored. An AP is typically composed of four elements, written A:B::C:D when the ratio between A and B is conform with the one between C and D, and this formal tool has been declined to cover numerous different interpretations of what an analogy can be, with any number of elements. In our work, we explore different methods to tackle the detection of APs and solving of analogical equations on different domains, using Deep Learning (DL). In particular, we study word morphology and Target Sense Disambiguation for words in context, two domains for which our model outperform the State of the Art (SotA), as well as in frame semantics, where we obtain encouraging results. We also study CoAT, a Case-Based Reasoning (CBR) system based on analogical transfer. With CoAT, we encounter significant success in the measure of case competence and in the task of case base compression.
Invited guests :
