RT @EPFL_en: Researchers from our Geometric Computing Laboratory have developed a new way to design complex, curved three-dimensional shape….

Building a karaoke booth at EPFL? Yup, you read that right! 🎤.ICONS, Episode 1 – where IC students do more than just code. 🌟 . Get the inside scoop on this creative project with Aya Hamane and Roxane Chevalley. #icons #epfl #student #icepfl

Researchers at EPFL have found that humans still outperform AI at recognizing objects with missing or fragmented outlines. Their study shows how training AI on more human-like visual data could help close the gap. 🔗Read more:

Researchers at EPFL have created a mathematical model that helps explain how breaking language into sequences makes modern AI like chatbots so good at understanding and using words. 🔗 Read more:

RT @zamir_ar: We open-sourced the codebase of Flextok. Flextok is an image tokenizer that produces flexible-length token sequences and repr….

RT @EPFL_en: EPFL and @CHUVLausanne are part of an African-EU partnership to roll-out an AI-driven application for more accessible and cost….

RT @EPFL_en: 🎉 Congratulations to our researchers who were selected by the @ERC_Research as part of the 2024 call for proposals for the Adv….

RT @EPFL_en: Can Switzerland take its place in GenAI? With specific entities created to facilitate collaboration and support the country’s….

RT @EPFL_en: ➕ 🔡 Our researchers have created a mathematical model that helps explain how breaking language into sequences makes modern AI….

In the power-law case, he finds that although individual neurons are learned suddenly and at different timescales, their overlapping of emergent learning curves across timescales create a smooth scaling law in the cumulative objective.💡📈.

He characterizes the dynamics of SGD and identifies sharp transition times when the network begins to recover individual signal directions—highlighting abrupt, phase-transition-like behavior in learning.

He investigates how online stochastic gradient descent (SGD) learns a two-layer neural network with many orthogonal neurons (M ≫ 1) on isotropic Gaussian data—even under strongly varying second-layer weights, including power-law scaling.

Denny Wu is a Faculty Fellow at NYU’s Center for Data Science & Flatiron Institute’s Center for Computational Mathematics. Previously, he completed his PhD at University of Toronto & the Vector Institute (advised by Jimmy Ba & Murat Erdogdu), and undergrad at CMU. 🎓.

Join us today for an @EPFL_AI_Center seminar: Learning Shallow Neural Networks in High Dimensions — SGD Dynamics and Scaling Laws.🎙️ Speaker: Denny Wu.🤝 Co-organized with the EPFL Foundations of Learning and AI Research (FLAIR) group. #epfl #aicenter #research.

🎉 Congratulations to Anastasia Ailamaki, Mathias Payer, Lenka Zdeborová & Pascal Fua for receiving Advanced Grants for research in data systems, security, neural networks & computer-assisted engineering, a record 4 awards for EPFL’s IC School! 🏆

Why are symbols used this way?. As cooperation is based on trust, but with social networks come with risks of polarization and fake news, but also privacy and activism, semantics were involved into understanding those social cues and patterns.

First, how to manage symbols efficiently?. Several solutions were tested such as distributed hash tables, which would lead to some advantages and disadvantages. However Aberer’s lab rigorously tackled P-Grids and were subject to a lot of PhDs thesi, which he credited.

This type of programming answers us the questions we could ask about semantics organization, which will be successively tackled.

Prof. Aberer also tackles the differences between computation, delving into fix-point programming. In fact, fix-point programming gives room for semantics with modularity, distributability, parallelism and tolerance to incompleteness👾✨

First, he begins by stating what semantics are, being about understanding the meaning of symbols and their relations. Semantics are context dependent and raise many questions computation-wise!