Software Institute

In February 2019, the Software Institute started its SI Seminar Series. Every Thursday afternoon, a researcher of the Institute will publicly give a short talk on a software engineering argument of her choice. Examples include, but are not limited to, novel interesting papers, seminal papers, personal research overview, discussion of preliminary research ideas, tutorials, and small experiments.

On our YouTube playlist you can watch some of the past seminars. Below you can find more details on the next seminar, the upcoming seminars, and an archive of the past speakers.

Everyone is welcome to attend the seminars organized by the Software Institute.

Next Speaker: Gianmarco De Vita

Date: November 21, 2024 @ 16:30

Location: D1.13

Chair: Marco Paganoni

Topographical Deep Learning Testing

Disclaimer: The content of this seminar has been already covered in the presentation of my prospectus.
The ubiquity of Deep Learning (DL) software in various domains has triggered a substantial amount of research work related to testing of DL systems. The main challenges with DL testing lie in the nature of the inputs and of the models. In fact, the input is complex, often high dimensional, and needs proper pre-processing to reduce its dimensionality and to identify the features that can discriminate a correctly behaving model from an incorrectly behaving one. For that purpose, this research work focuses on addressing the testing procedure from a model-agnostic perspective. Specifically, we want to characterise the input space topographically for any possible model—including different model architectures—that operate on a given domain. Within the input space, we are interested in identifying the features shared by the inputs that are more likely to trigger misbehaviour in the tested DL systems. The only DL testing approach targeting explainable features, DeepHyperion, requires a deep domain knowledge as well as extensive manual effort to define domain-specific features. Instead, our approach is designed to automate the partitioning of the feature space into regions and identifying the areas where inputs share fault-revealing features. The aim of this research is to address DL testing from a topographical perspective: (i) generation of a topographical map of the input space through an automated procedure that can be generalized to diverse domains; (ii) investigation on whether the features identified in the map correlate with mutation killing capabilities of inputs; (iii) assessment of a proper prioritisation of inputs based on their features to reduce the costs of manual input labelling; (iv) input generation guided by fault-inducing features extracted from the topographical map; (v) input generation through the interpolation between existing inputs.

Biography

Gianmarco De Vita is a Ph.D. candidate in the Faculty of Informatics at the Università della Svizzera italiana (USI), Switzerland, where he is part of the TAU Research Group. He received his Master’s degree in Informatics at USI in 2023. His current research concentrates on testing of deep learning systems and exploration of their input space.

Program

Rosalia Tufano

November 28, 2024
Akshatha Shenoy

December 5, 2024
Dominik Winterer

February 27, 2025

Seminars

Next Speaker: Gianmarco De Vita

Program

Archive