Date/Time: Monday, November 27, 12:00 – 13:30
Location: FORTH Central Building - Payatakis Room
Speaker: Prof. Alexandros Iosifidis, Aarhus University, Denmark
Title: Continual Inference Networks for Real-time Stream Data Processing
Hosts: Dr. George Tzagkarakis, SPL LAB / ICS FORTH
Abstract Most of the state-of-the-art Deep Neural Network architectures processing sequence data (e.g., videos) like 3D Convolutional Neural Networks (3D-CNNs) and Transformers were built for offline processing. This means that, instead of processing newly captured input data one at a time, they require the entire sequence to be passed as a single input. Yet, many important real-life applications, such as tasks in Robot Perception and visual monitoring in Smart Cities, need online predictions on a continual input (visual) streams. While 3D-CNNs and Transformers can be applied by re-assembling and passing sequences within a sliding window, this is inefficient due to the redundant intermediary computations from overlapping clips. In this talk, I will present our recently introduced Continual Inference Networks (CINs) which are built to ensure efficient stream data processing. This is achieved by employing an alternative computational ordering that allows sequential computations without the use of sliding window processing. Concrete realizations of CINs are the Continual 3D-CNNs, the Continual Spatio-Temporal Graph Convolutional Networks (ST-GCNs) and the Continual Transformers, which require approximately L x fewer number of computations per prediction compared to sliding window-based inference with non-CINs, where L is the corresponding sequence length of a non-CIN network, while not sacrificing in accuracy.
Short bio: Alexandros Iosifidis is a Professor at Aarhus University, Denmark. He is the director of the Machine Learning & Computational Intelligence group at the Department of Electrical and Computer Engineering, and he is leading the Machine Intelligence research area of the University's Centre for Digitalisation, Big Data and Data Analytics (DIGIT). He has contributed to more than thirty R&D projects financed by EU, Finnish, and Danish funding agencies and companies. He has (co-)authored 110+ articles in international journals and 150+ papers in international conferences/workshops in topics of his expertise. He is a co-Editor of the Deep Learning for Robot Perception and Cognition book (Academic Press, 2022). Alexandros is currently serving as the Associate Editor in Chief for the Neurocomputing journal covering the research area of Neural Networks. He has been serving in the Editorial Board of international journals, including the IEEE Transactions on Neural Networks and Learning Systems and the IEEE Transactions on Artificial Intelligence. He contributed to the organization of international conferences as an Area Chair or Technical Program Committee Chair, including IEEE ICIP (2018-2023), IEEE ICASSP (2023, 2024) and EUSIPCO (2019,2021,2023), and as Publicity co-Chair for IEEE ICME 2021.