Pervasive AI: (deep) Learning into the Wild
Davide Bacciu, University of Pisa, Italy
Deep Reinforcement Learning to Improve Traditional Supervised Learning Methodologies
Luís Paulo Reis, University of Porto, Portugal
Pervasive AI: (deep) Learning into the Wild
Davide Bacciu
University of Pisa
Italy
http://pages.di.unipi.it/bacciu
Brief Bio
Davide Bacciu is Associate Professor at the Department of Computer Science, University of Pisa, where he is the founder and head of the Pervasive Artificial Intelligence Laboratory. He holds a Ph.D. in Computer Science and Engineering from the IMT Lucca Institute for Advanced Studies, for which he has been awarded the 2009 E.R. Caianiello prize for the best Italian Ph.D. thesis on neural networks. He has co-authored over 160 research works on (deep) neural networks, generative learning, Bayesian models, learning for graphs, continual learning, and distributed and embedded learning systems. He is the coordinator of two EC-funded projects on embedded, distributed and lifelong learning AI, and of several national/industrial projects. He is the chair of the IEEE CIS Neural Network Technical Committee and a Vice President of the Italian Association for AI.
Abstract
The deployment of intelligent applications in real-world settings poses significant challenges which span from the AI methodology itself to the computing, communication and orchestration support needed to execute it. Such challenges can inspire compelling research questions which can drive and foster novel developments and research directions within the deep learning field. The talk will explore some current research directions in the context of pervasive deep learning, touching upon efficient neural networks, non-dissipative neural propagation, neural computing on dynamical systems and continual learning.
Deep Reinforcement Learning to Improve Traditional Supervised Learning Methodologies
Brief Bio
Luis Paulo Reis is an Associate Professor with Habilitation at the University of Porto in Portugal and Director of LIACC – Artificial Intelligence and Computer Science Laboratory. He is an IEEE Senior Member and he was president of the Portuguese Society for Robotics and of the Portuguese Association for Artificial Intelligence. He is Co-Director of LIACD - First Degree in Artificial Intelligence and Data Science. During the last 25 years, he has lectured courses, at the University, on Artificial Intelligence, Intelligent Robotics, Multi-Agent Systems, Simulation and Modelling, Games and Interaction, Educational/Serious Games and Computer Programming. He was the principal investigator of more than 20 research projects in those areas. He won more than 60 scientific awards including winning more than 15 RoboCup international competitions (including the last 3 editions of the Simulation 3D League - Humanoid Robots) and best papers at conferences such as ICEIS, Robotica, IEEE ICARSC and ICAART. He supervised 24 PhD and 160 MSc theses to completion and is supervising 12 PhD theses. He evaluated more than 50 projects and proposals for FP6, FP7, Horizon2020, FCT, and ANI. He was a plenary speaker at several international conferences such as ICAART, ICINCO, LARS/SBR, WAF, IcSports, SYROCO, CLAWAR, WCQR, ECIAIR, DATA/DELTA and IC3K. He organized more than 70 international scientific events and belonged to the Program Committee of more than 300 scientific events. He is the author of more than 450 publications in international conferences and journals (indexed at SCOPUS or Web of Knowledge).
Abstract
This talk focuses on the intersection of Deep Reinforcement Learning (DRL) and traditional Supervised Learning (SL) methodologies, exploring how DRL can enhance performance and overcome challenges in tasks typically approached via SL. Despite the success of SL in various domains, its limitations, including the inability to handle sequential decision-making and non-stationary environments, are obvious, making DRL a potentially useful tool.
The talk will outline the fundamental principles of DRL, including its distinguishing features, such as learning from delayed rewards, handling the exploration-exploitation trade-off, and operating in complex, dynamic environments. It will also focus on the integration of DRL into traditionally SL-dominated areas, providing real-world examples from several fields. The talk will discuss how DRL can automate and optimise processes within the machine learning pipeline that have traditionally been manual and heuristic, such as hyperparameter tuning and feature engineering. By using DRL, the talk will showcase how these processes can be transformed into learnable tasks, improving the efficiency and performance of the supervised learning system. The talk will also present the latest research and techniques on the incorporation of DRL into traditionally SL-focused domains and feature interesting examples from several projects developed at the University of Porto on these areas of DRL and DRL for SL, such as the DRL methodologies included in our RoboCup world champion team in the humanoid 3D Simulation League.