Some recent advancements in machine learning for embedded systems

DISI Seminars

January 26, 2021
Versione stampabile

Online event on Zoom, 10:30 am

Speaker

  • Claudio Turchetti, Università Politecnica delle Marche

Abstract    

Machine learning (ML) techniques are both computationally and memory intensive, making them difficult to deploy on embedded systems with limited hardware resources. In many practical applications the dimensionality of input vectors is very high, so many machine learning techniques incur the “curse of dimensionality” issue, or require depth neural networks (DNN/CNN) with a large number of weights that consume considerable storage and memory bandwidth. Dimensionality reduction of data is a powerful technique that aims to transform high-dimensional data into a desired low-dimensional representation, thus reducing the computational cost of ML applications. The aim of this talk is twofold, first we will present some recent approaches for dimensionality reduction of data with particular emphasis to Manifold Learning , secondly techniques for compression of Convolution Neural Network (CNN), to allow their usage in resource-limited embedded devices, will be discussed.

Keywords: Machine Learning, Nonlinear Dimensionality Reduction, Intrinsic Dimension, Manifold Learning, CNN Compression

Event dedicated to: PhD students, researchers in Electrical and Electronics Engineering, Computer Science, Information Sciences. Participants are invited to ask questions via chat.

Registration

Participation in this event is free, but due to organizational reason you are invited to register on line
The Zoom link is now available.

About the speaker 

Claudio Turchetti received the Laurea degree in electronics engineering from the University of Ancona, Ancona, Italy, in 1979. He joined the Università Politecnica delle Marche, Ancona, in 1980, where was the Head of the Department of Electronics, Artificial Intelligence and Telecommunications for five years and is currently a Full Professor of Micro-Nanoelectronics and Embedded Systems.
His current research interests include: statistical device modeling, RF integrated circuits, device modeling at nanoscale, computational intelligence, signal processing, pattern recognition, system identification, machine learning and neural networks.
He has published more than 160 journal and conference papers, and two books.
The most relevant papers were published in IEEE J. of Solid-State Circuits, IEEE Trans. on Electron Devices, IEEE Trans. on CAD of IC’s and Systems, IEEE Trans. on Neural Networks and Learning Systems, IEEE Trans. on Signal Processing, IEEE Trans. on Cybernetics, IEEE J. of Biomedical and Health Informatics, IEEE Trans. on Consumer Electronics, Information Sciences.
He has held a variety of positions as Project Leader in several applied research programs developed in cooperation with small, large, and multinational companies in the field of microelectronics. At present is involved in three European Projects. Prof. Turchetti has served as a Program Committee Member for several conferences and as a reviewer of several scientific journals. He is a Member of the IEEE, Computational Intelligence and Signal processing Society. He has been an Expert Consultant of the Ministero dell’Università e Ricerca.

DISI Seminars are meant as the primary venue to discuss scientific and research issues that are of interest to the Department as a whole. Our students, researchers and academics will have the opportunity to network and to share knowledge and experience with renowned invited speakers, who lead their research at the highest standards through a multi- and interdisciplinary approach. 
One of the objectives of this series is to foster interaction across different research areas, in a horizontal perspective ranging from engineering to computer science. We strongly believe that cooperation and interdisciplinary skills are fundamental pillars on which to lay the foundations of successful research.