“How Neural Networks Simulate and Learn Algorithms”, Kimon Fountoulakis, Associate Professor, David R. Cheriton School of Computer Science, University of Waterloo, Canada – Τμήμα Μηχανικών Η/Υ και Πληροφορικής

Την ερχόμενη Παρασκευή 23/5 θα έχουμε τη μεγάλη χαρά να φιλοξενήσουμε “εξ αποστάσεως” ως ομιλητή στα πλαίσια των εκδηλώσεων “Σεμινάριο CEID & Social Hour” και των ΔΠΜΣ ΥΔΑ, ΣΜΗΝ τον δρ. Κίμωνα Φουντουλάκη, Associate Professor, David R. Cheriton School of Computer Science, University of Waterloo. Όπως πάντα, θα ακολουθήσει συζήτηση και θα προσφερθεί καφές.

Please note the following interesting and highly topical talk that will be presented in the context of the weekly event “CEID Seminar & Social Hour” organized by CEID, and the MS programs DDCDM, SMIN

Τίτλος:  How Neural Networks Simulate and Learn Algorithms

Ομιλητής: Kimon Fountoulakis (εξ αποστάσεως), Associate Professor, David R. Cheriton School of Computer Science, University of Waterloo, Canada

Ημερομηνία-χώρος: Παρασκευή 23 Mαΐου,  3:15-5μμ, ΤΜΗΥΠ, αμφιθέατρο Γ

Abstract: Neural networks have recently demonstrated tremendous success in language, vision, and various reasoning tasks. However, they still perform poorly on tasks that require executing long algorithmic instructions, a form of reasoning known as algorithmic reasoning. This type of reasoning lies at the core of fundamental tasks like arithmetic and sorting, which are often used as testbeds for evaluating algorithmic execution in neural networks. In this talk, I will discuss, from a theoretical perspective, the ability of various architectures, such as feedforward networks, Transformers, and graph neural networks, to simulate and learn, through training, how to execute algorithms and general computer instructions. Finally, I will discuss their ensemble complexity, that is a lower bound on the number of training trials required for a neural network to learn to execute algorithms with zero error.