BEGIN:VCALENDAR VERSION:2.0 PRODID:-//Date iCal//NONSGML kigkonsult.se iCalcreator 2.20.2// METHOD:PUBLISH X-WR-CALNAME;VALUE=TEXT:Eventi DIAG BEGIN:VTIMEZONE TZID:Europe/Paris BEGIN:STANDARD DTSTART:20251026T030000 TZOFFSETFROM:+0200 TZOFFSETTO:+0100 RDATE:20261025T030000 TZNAME:CET END:STANDARD BEGIN:DAYLIGHT DTSTART:20260329T020000 TZOFFSETFROM:+0100 TZOFFSETTO:+0200 TZNAME:CEST END:DAYLIGHT END:VTIMEZONE BEGIN:VEVENT UID:calendar.30778.field_data.0@www.diag.uniroma1.it DTSTAMP:20260502T094229Z CREATED:20260428T132128Z DESCRIPTION:AbstractCurrent AI architectures remain significantly inferior to humans and animals in their ability to reason\, plan\, and understand t he physical world. This lecture unpacks the rationale for moving beyond su pervised and reinforcement learning toward objective-driven AI. We will ex plore Moravec’s paradox — the discrepancy between human data efficiency an d computational difficulty — and the inherent limitations of autoregressiv e models that lead to hallucinations and compounding errors. The talk focu ses on the transition from simple feedforward prediction (‘system 1’) towa rd world models capable of deliberate reasoning and inference through opti misation (‘system 2’). We will introduce the conceptual framework of Joint -Embedding Predictive Architectures (JEPA) and Energy-Based Models (EBMs) as a unifying mathematical language for self-supervised learning.In the se cond half of the lecture\, the audience will be invited to choose one of t wo directions for a practical deep-dive:Option A: recursion and optimal co ntrol. This path examines the ‘action’ component of the world model. We wi ll discuss the role of recursion as embodied by Recurrent Neural Network ( RNN) equations and show how to use backpropagation through time to find op timal sequences of actions via Model-Predictive Control (MPC).Option B: la tent-variable energy-based models. This path uses a toy architectural exam ple to master the mechanics of EBMs. We will define the relationship betwe en energy and free energy\, explore various loss functionals\, and demonst rate how inference is performed as a minimisation problemUltimately\, this lecture provides a technical roadmap for moving beyond Generative AI towa rd autonomous and controllable systems. It addresses the why by analysing the data inefficiency of current paradigms\, the what by defining objectiv e-driven world model architectures\, and the how by exploring the mathemat ical foundations of Joint-Embedding Predictive Architectures (JEPA) and En ergy-Based Models (EBMs).Speaker reference: https://atcold.github.io DTSTART;TZID=Europe/Paris:20260521T150000 DTEND;TZID=Europe/Paris:20260521T150000 LAST-MODIFIED:20260428T151722Z LOCATION:Aula B2 - DIAG (In Remote) SUMMARY:Self-Supervised Learning\, JEPA\, World Models\, and AI’s future - Alfredo Canziani\n\n\n \n \n\n \n\n\nIrene\n\n\nAmerini \n\n \n\n \n\n\n\n\n\nProfessore associato\n\n\npagina personale\n\nstanza: \n\nB2 15\n\ntelefono: \n\n+39 0677274044 \n\n \n\n \n\nBiografia: \n\n\n\nI rene Amerini is Associate Professor at the Department of Computer\, Contro l and Management Engineering A. Ruberti of Sapienza University of Rome\, I taly where she is leading the Computer VIsion and Multimedia Forensics Res earch Team at ALCORLab.\n\n\nFrom 2019-2022\, she was an Assistant Profess or at DIAG\, Sapienza University of Rome and previously a postdoctoral res earcher at the Image Forensics and Security Lab\, Media Integration and Co mmunication Center\, University of Florence (Italy). In 2018 she obtained a Visiting Research Fellowship at Charles Sturt University (Australia) off ered by the Australian Government – Department of Education and Training t hrough the Endeavour Scholarship & Fellowship program and the Italian Habi litation for Associate Professor in Computer Science and Telecommunication s.\n\n\nIn 2010 she spent part of her PhD course at the Digital Data Embed ding Laboratory\, Department of Electrical and Computer Engineering\, Bing hamton University (US). In 2011 she received the Ph.D. in computer enginee ring\, multimedia and telecommunication from the University of Florence (I taly) with the thesis “Image forensics: source identification and tamperin g detection”.\n\n\nIn 2019 she co-organized the Woman In Computer Vision W orkshop (WiCV) at CVPR 2019 in Long Beach\, California\, in 2020 the Woman In Computer Vision Workshop (WiCV) at ECCV\, the W4PR (Woman at ICPR Work shop) in conjunction with ICPR 2021 and of the of the VERIMEDIA workshop i n conjuction with IJCNN July 2025.\n\n\nShe is currently a member of the I EEE Forensics and Security Technical Committee (IFS-TC)\, EURASIP TAC Biom etrics\, Data Forensics\, and Security and IAPR TC6 - Computational Forens ics Committee and Senior Area Editor (S-AE) for the SPS Signal Processing Letters (SPL) May 2025-2027.\n\n\nHer main research activities include com puter vision\, pattern recognition\, adversarial machine learning and mult imedia forensics.\n\n\n \n\n\nList of publications on dblp or Google schol ar\n\n\nEmail: amerini AT diag.uniroma1.it\n\n\nWebsite/Teaching\n\n\nInte ressi di ricerca: \n\nMultimedia forensics and deepfake analysis\nMachine learning and deep learning for image and video analysis\nDigital image and video processing\nAdversarial machine learning/deep learning\nComputer vi sion and monocular depth estimation\n\n\nqualifica_rr: \n\nAssociate profe ssors \n\n \n\n \n\n\n\n\, \n\n\n \n \n\n \n\n\nGianmarco\n\n \nScarano \n\n \n\n \n\n\n\n\n\ndottorando\n\nMember of: \n\n \n\n \n\n \n\n \n\n \n\n\n\n\, \n\n\n \n \n\n \n\n\nClaudio\n\n \nSchiavella \n\n \n\n \n\n\n\n\n\ndottorando\n\n\npagina personale\n \nstanza: \n\nS012\n\ntelefono: \n\n35155\n\nMember of: \n\n \n\n \n\n \n\nBiografia: \n\n\n\nSono un dottorando in Ingegneria Informatica pres so la Sapienza Università di Roma (pagina personale PhD)\, attualmente imp egnato in attività di ricerca presso l'ALCOR Computer Vision Lab. Il mio l avoro si concentra sullo sviluppo e l'ottimizzazione di modelli innovativi e portabili di Computer Vision\, progettati per dispositivi edge\, con l' obiettivo di spingere i limiti di efficienza e flessibilità di implementaz ione in applicazioni reali.\n\n\nHo conseguito una laurea triennale in Ing egneria Informatica e dell'Automazione e una laurea magistrale in Intellig enza Artificiale e Robotica\, laureandomi con il massimo dei voti. Il mio percorso accademico mi ha fornito solide basi in calcolo avanzato\, sistem i intelligenti e automazione\, che guidano la mia attuale ricerca nella co mputer vision all'avanguardia per dispositivi edge.Ho maturato esperienza internazionale studiando presso diverse università e scuole in Europa e As ia\, tra cui Repubblica Ceca\, Francia\, Portogallo\, Finlandia\, Inghilte rra\, Scozia\, Irlanda e Giappone. Sono appassionato di startup innovative nell'ambito dell'Industria 4.0 e profondamente impegnato nell'insegnament o. Ho esperienza come docente nelle scuole superiori\, dove mi piace coinv olgere gli studenti attraverso un approccio pratico e interattivo all'appr endimento.\n\n\n \n\n\nInteressi di ricerca: \n\nModelli leggeri per la Co mputer Vision\nDeep learning geometrico per la Computer Vision\nConformal prediction per la Computer Vision\n\n\nkeywords: \n\nArtificial Intelligen ce and Robotics URL;TYPE=URI:https://www.diag.uniroma1.it/node/30778 END:VEVENT END:VCALENDAR