Date: Thursday, 15th October 2020
Location: TEAMS - code to join the team: n25vdtt
Speaker: Dario Guidotti
Time: 2.30 pm
Title: Verification and Repair of Neural Networks
Abstract:
Neural Networks are becoming popular in many technological domains. Many such domains are, in some way, safety or security-critical: as such, besides good performance in the tasks of interests, they require formal guarantees on the behaviour of the systems involved. Whereas systems based on neural networks are able, in general, to satisfy the performance requirements they do not provide any formal guarantee on their behaviour. In this seminar, we show some research results on the enhancement of the scalability of existing verification methodologies and on the repair of neural networks.
Bio:
I am a DPhil student in Computer Science and System Engineering under the supervision of Armando Tacchella and Massimo Narizzano (DIBRIS – UNIGE). My main field of research is the enhancement of machine-learned models through formal verification and, in particular, the verification of neural networks. To be more specific the aim of my work is to show, using formal methods, that some properties of interest hold for a certain machine learning model and, if necessary, to repair such model in order to guarantee the validity of the properties.