Deep Learning at the Edge
Join the event in Linz
ALOHA is an H2020 project, started in January 2018, on a software framework for runtime-adaptive and secure deep learning on embedded systems.
The event within H2020 ALOHA project is dedicated towards topics covered in the project development, such as Deep learning, Embedded systems, Security of Deep learning architectures. Speakers from Italy, Switzerland and Austria will share their expert knowledge and provide information about ALOHA tool development.
15:00 - 15:10 Introduction
15:10 - 15:50 Deep Learning: Research & Applications
Bernhard Nessler, Johannes Kepler University Linz (JKU), Austria
Bernhard Moser, Software Competence Center Hagenberg (SCCH), Austria
We give an overview of running and planned research projects at JKU and SCCH on deep learning with applications in various fields such as mobility, surveillance, industry and bioinformatics. In this context we illustrate potentials and challenges of emerging developments.
15:50 - 16:30 Cognitiveness at the edge: Platforms, Models, Tools -an insight into
the ALOHA project
Paolo Meloni, University of Cagliari (UniCa), Italy
We present challenges related with bringing cognitive intelligence to edge CPS devices, we discuss the state of the art, focusing on novel processing platforms and on utilities supporting designers and programmers. We also give an overview of the current status of the ALOHA H2020 research project, focusing on efficient and secure running of Deep Learning algorithms at the edge.
16:30 - 16:50 Coffe break
16:50 - 17:30 Bringing Deep Learning to the Edge
Francesco Conti, Swiss Federal Institute of Technology in Zurich (ETH Zurich), Switzerland
Deep Learning and Deep Neural Networks (DNNs) have emerged in the last few years as the go-to algorithmic choice for any application that requires advanced artificial intelligence capability. The high workload and energy cost of DNNs, however, have so far hindered their application to devices such as IoT nodes and cyber-physical systems that have to operate under stringent constraints.
In this talk, we will focus on techniques that can be used to actually bring DNNs to the edge and to the real world for applications such as autonomous UAV guidance, using a combination of specialized hardware and controlled algorithmic approximations.
17:30 - 18:10 Evaluating Security of Deep Learning to Adversarial Examples
Maura Pintor, Pluribus One, Italy
Deep learning has obtained impressive results in many tasks, from computer vision to speech recognition, thanks to the increasing availability of data, hardware and software tools, raising the attention of the scientific and industrial communities, and of society at large. However, it has been shown that such systems can be misled by adversarial examples, i.e., opportunely-modified input data that cause these algorithms to fail their main task of understanding what the input represents. Depending on the application, the risk of an attack causing great damage can be high. In this talk, I will discuss some attack algorithms capable of generating adversarial examples, how to use them to evaluate the robustness of a deep network, and how such threats can be countered and mitigated, in the context of specific application examples. To this end, I will also show a concrete demonstration of the security evaluation tool that we developed in the context of the H2020 ALOHA project.
18:10 - 19:00 Poster Session