1. Speaker: Tomasz Trzciński (Warsaw University of Technology) -
Title: Deep generative view on continual learning
Course summary: Neural networks suffer from catastrophic forgetting, defined as an abrupt performance loss on previously learned tasks when acquiring new knowledge. For instance, if a network previously trained for detecting virus infections is now retrained with data describing a recently discovered strain, the diagnostic precision for all previous ones drops significantly. To mitigate that, we can retrain the network on a joint dataset from scratch, yet it is often infeasible due to the size of the data, or impractical when retraining requires more time than it takes to discover a new strain. The catastrophic forgetting severely limits the capabilities of contemporary neural networks and continual learning aims to address this pitfall. During the course, we will introduce continual learning as a domain of machine learning, define its main challenges and existing methods. We will then look at the approaches inspired by recent neuroscientific works, specifically at the generative models employed in continual learning scenario. Based on our research where we hypothesize that the unsupervised way of incorporating knowledge by generative models corresponds to the way biological systems continually learn, we will introduce generative replay as a method to overcome catastrophic forgetting. We will then further explore the landscape of continual learning from a deep generative modeling point of view.
Bio: Tomasz Trzciński (DSc, WUT'20; PhD, EPFL'14; MSc, UPC/PoliTo'10) is an Associate Professor at Warsaw University of Technology since 2015, where he leads a Computer Vision Lab, and an Research Group Leader at IDEAS NCBR, a publicly-funded AI center in Poland. He was a Visiting Scholar at Stanford University in 2017 and at Nanyang Technological University in 2019 and 2023. Previously, he worked at Google in 2013, Qualcomm in 2012 and Telefónica in 2010. He is an Associate Editor of IEEE Access and MDPI Electronics and frequently serves as a reviewer in major computer science conferences (CVPR, ICCV, ECCV, NeurIPS, ICML) and journals (TPAMI, IJCV, CVIU). He is a Senior Member of IEEE, ELLIS Member and director of ELLIS Unit Warsaw, and an expert of National Science Centre and Foundation for Polish Science. He is a Chief Scientist at Tooploox and a co-founder of Comixify, a technology startup focused on using machine learning algorithms for video editing.
2. Speaker: Simone Scardapane (Sapienza University of Rome), Gabriele Tolomei (Sapienza University of Rome) - email@example.com, firstname.lastname@example.org
Title: Explainable and interpretable AI
Date: May 2024
Abstract: The course will introduce basic concepts related to explaining and debugging machine learning and artificial intelligence models. In particular, we will cover feature attribution methods, data attribution methods, and counterfactual explanations. The course will have both theory and practical sessions, as long as possible guest speakers.
Bio: Gabriele Tolomei is an associate professor at the Department of Computer Science of the Sapienza University of Rome. He received his M.Sc. in Computer Science in 2005 from the University of Pisa and his Ph.D. in Computer Science in 2011 from the Ca' Foscari University of Venice. He was a research scientist at Yahoo Labs in London and an assistant professor at the Department of Mathematics at the University of Padua. His main research interests include (human-)explainable, robust, and collaborative machine learning. He authored more than 40 papers in peer-reviewed international conferences and journals, and he is the inventor of four US patents.
3. Speaker: Franco Maria Nardini (STI-CNR) - email@example.com
Title: Challenges in Modern Web Search
Date: September/October 2024
Abstract: This PhD course focuses on Web search and discusses the challenges in the three main areas of Web search: i) crawling, ii) indexing, and iii) query processing. The course introduces each area by discussing the state of the art in the field and by presenting the open research questions. The course emphasizes query processing, an area where machine learning is important to advance the state of the art. After introducing the different query processing techniques, the course i) introduces supervised techniques explicitly focused on targeting the ranking problem, ii) discusses several efficiency/effectiveness trade-offs in query processing, and iii) analyze several related optimization techniques. The course will also provide an overview of the role of pre-trained large language models in query processing techniques. Two hands-on sessions will cover indexing and query processing of public Web collections.
Bio: Franco Maria Nardini is a Senior Researcher with ISTI-CNR in Pisa, Italy. His research interests are focused on Web Information Retrieval, Machine Learning, and Data Mining. He authored over 100 papers in peer-reviewed international journals, conferences, and other venues. In the past, he has been Program Committee Co-Chair of SPIRE 2023, Tutorial Co-Chair of ACM WSDM 2021, Demo Papers Co-Chair of ECIR 2021, Program Committee Co-Chair, and General Co-Chair of ReNeuIR at SIGIR (2022, 2023). He is a co-recipient of the ECIR 2022 Industry Impact Award, the ACM SIGIR 2015 Best Paper Award, and the ECIR 2014 Best Demo Paper Award. He is a member of the editorial board of ACM TOIS and a program committee member of SIGIR, ECIR, SIGKDD, CIKM, WSDM, IJCAI, and ECML-PKDD.
4. Speaker: Guillem Rigail (INRAE) - firstname.lastname@example.org
Title: An introduction to (Multiple) Changepoint detection
Abstract: In recent years, there has been a proliferation of methods for detecting changepoints (also known as breakpoints or structural breaks) in data streams. This surge has been driven by the wide range of applications where changepoint methods are needed, including genomics, neuroscience, climate science, finance, and econometrics, among others. This course serves as an introduction to multiple changepoint detection methods.
Initially, I will address the simpler task of detecting a single changepoint in the mean of a univariate data stream. This is crucial for understanding several state-of-the-art approaches designed for detecting multiple changepoints. Subsequently, I will delve into the fundamentals of two classical approaches for multiple changepoint detection: (1) binary segmentation and (2) dynamic programming. I will review their statistical and computational properties and explain some of their recent improvements.
I will illustrate the application of these approaches to genomic datasets using the R programming language.
Bio: Guillem Rigaill is senior researcher (aka DR) at INRAE in France. He is a member of the GNet Team at the Institute of Plant Sciences Paris-Saclay and the Stat & Genome team at the Laboratoire de Mathématiques et Modélisation d'Évry. He received his PhD from AgroParisTech in 2010 for the development of algorithms and statistical methods for the analysis of breast cancer data.
His research interests focus on the development of efficient algorithms and appropriate statistical methodologies for the analysis of high-dimensional genomic and transcriptomic data. He has been developing new models for changepoint detection and proposed inference procedures for these models that are both statistically and algorithmically efficient. He has applied those new tools in a number interdisciplinary projects involving cancer and plant biologists, bioinformaticians and statisticians.