You are here

Neural-Network Interpretability for Time Series Classification Task

Gianmarco Mengaldo
Tuesday, July 18, 2023 - 16:00 to 17:00
A-133 and online

Neural networks (NN) have been gaining significant traction for time series classification tasks over the past few years. Yet, they are frequently perceived as black-box tools, whose results may be difficult to interpret. To address this issue, several methods have been proposed to obtain maps of relevance scores highlighting the importance of different time steps for a given model. These methods were initially applied to images, and more recently to time-series data. Yet, interpretability of NN remains challenging. Indeed, interpretability methods typically provide different results, sometimes even diametrically opposite, and may not explain how neurons collaborate to represent specific patterns. In this work, we propose a new evaluation framework for post-hoc interpretability methods applied to time series classification tasks. We argue that this work is a critical step toward understanding NN-based decisions and provide a more robust interpretability workflow. We also present a preliminary study that aims to understand the robustness of the evaluation metrics.

Dr Gianmarco Mengaldo is an assistant professor in the Department of Mechanical Engineering at National University of Singapore (Singapore), and Honorary Research Fellow at Imperial College London (United Kingdom). He received his BSc and MSc in Aerospace and Aeronautical Engineering from Politecnico di Milano (Italy), and his PhD in Aeronautical Engineering from Imperial College London (United Kingdom). After his PhD he undertook various roles both in industry and academia, including at the European Centre for Medium-Range Weather Forecasts, the California Institute of Technology, and Keefe, Bruyette and Woods (KBW). Dr Mengaldo’s adopts an interdisciplinary approach integrating mathematical and computational engineering to study complex systems that arise in applied science. His current research interests involve (i) the development of data-mining technologies for the systematic identification of coherent patterns in highly unstructured datasets, (ii) the development of high-fidelity simulations tools for multi-physics problems, and (iii) the use of machine learning and statistical tools to predict the behaviour of complex systems. Dr Mengaldo’s main application areas include aerospace and mechanical engineering, weather and climate, solid earth physics, healthcare, and finance.

Sign in