A random solution to efficient large scale machine learning



Lorenzo Rosasco

A random solution to efficient large scale machine learning
Speaker:
Lorenzo Rosasco
Abstract:
The world is awash with data, and the corresponding huge potential source of information has been called a data revolution. At the root of this revolution is machine learning: a branch of artificial intelligence (AI) that studies and develops systems trained on data rather than being solely programmed. While the prospects yield by ML are thrilling, a close look at the resources needed by state of the art solutions is worrisome. Software and hardware advances make it easy to follow a brute force approach to ML development, based on deploying more and more resources. However, this ML growth model poses challenges that endanger its potential benefits. In this talk we discuss an instance of a new approach to make large scale machine learning more efficient and sustainable by combining different algorithmic ideas, from statistics to optimization, from algorithms, to numerical linear algebra. In particular, we analyze and discuss the properties of a class of algorithms that can be seen as shallow neural networks with random weights, and show both in theory and in practice how they can achieve state of the art learning accuracy, with only a fraction of computational resources.
Date:
Monday, March 26th, 2018 - 2:30PM
Location:
DIMA, Valletta Puggia, room 710, 7th floor
Sei qui: Home Ricerca Seminari di ricerca A random solution to efficient large scale machine learning