Back to TUM-PREP Projects overview

Project Overview

Project Code: CIT 18

Project name:

Neural Model-based Test Generation of Deep Learning-based Systems

TUM Department:

CIT - Informatics

TUM Chair / Institute:

Informatics 4 - Assistant Professorship of Software Engineering for Data-intensive Applications

Research area:

Software Engineering

Student background:

Computer ScienceComputer Science/ Informatics

Further disciplines:

Planned project location:

Boltzmannstr. 3, Garching bei München

Project Supervisor - Contact Details


Title:

Dr.

Given name:

Andrea

Family name:

Stocco

E-mail:

andrea.stocco@tum.de

Phone:

+491751164563

Additional Project Supervisor - Contact Details


Title:

Given name:

Family name:

E-mail:

Phone:

Additional Project Supervisor - Contact Details


Title:

Given name:

Family name:

E-mail:

Phone:

Project Description


Project description:

In safety-critical applications like autonomous vehicles, Deep Neural Networks play a crucial role. Testing these complex software systems is imperative due to their significance. As these networks learn policies from raw data to labels, understanding their precise decision boundaries presents challenges. Current test generators employ search-based methods to approximate these boundaries, but they rely on specific input data models (e.g., Catmull-Rom splines for self-driving car tracks) that are not widely applicable.

This project seeks to rethink existing model-based test generators for Deep Neural Networks with the aim to develop innovative techniques that can make them applicable comprehensively across various tasks and inputs. Specifically, we will delve into the potential of emerging generative adversarial networks (GANs) to learn meaningful embeddings for intricate data (e.g., driving images). These embeddings should be amenable to manipulation through search-based methods to create boundary-aware inputs in latent spaces.

The student's tasks encompass:
- Exploring the capabilities of generative adversarial networks in learning semantically meaningful embeddings.
- Implementing a Proof-of-Concept using an existing codebase.
- Assessing the Proof-of-Concept's performance on established datasets.
- Compiling a comprehensive report detailing the findings.


Why should you choose this proposal:
- Engage in an interdisciplinary project, gaining insights into software engineering and deep learning principles that will enhance your CV's value
- Join a young and collaborative team, fostering an environment of dynamic growth
- Experience a fulfilling and enjoyable journey

Upon successful admission, a detailed outline of tasks and objectives will be provided. Additionally, the student will be equipped with a comprehensive list of relevant literature prior to the project's start, aimed at fostering a solid grasp of the current advancements in the field. As your supervisor, I will offer continuous guidance throughout the research project, collaborating closely to brainstorming innovative ideas and navigate any potential challenges that may arise.

Looking forward to receiving your application and working with you!

Working hours per week planned:

32

Prerequisites


Required study level minimum (at time of TUM PREP project start):

3 years of bachelor studies completed

Subject related:

software engineering; machine learning; deep learning

Other:

  • Keine Stichwörter