Towards safe recurrent neural networks via model checking its interpretable structure

Description

About the Project:

The Faculty of Science & Engineering at the University of Liverpool is inviting applications for a fully-funded PhD studentship. We look for bright and motivated graduates, with a preference for those with expertise in formal methods. The successful candidate will join this cutting-edge research project `Towards Safe Recurrent Neural Network via Model Checking its Interpretable Structure`.

Neural networks are becoming a popular tool for solving many real-world problems, and the ubiquity of which will soon allow complex automatic systems to drive on our roads, fly over our heads, move around us during our daily lives and work in our factories. However, validation and verification of neural networks, especially recurrent neural networks (RNN), due to their more complex internal structures and their processing of sequential inputs with a temporal semantics, is still limited and leaving vast sequential application domains at risk.

This project will advance the state-of-the-art in the development of safe RNN models by comparing, using and extending interpretable structures (e.g. finite state automata and Markov decision processes) of RNN models, thus safeguarding them with formal verification techniques, especially quantitative model checking techniques. We will, therefore, develop methods, algorithms and tools to achieve fully verifiable RNN models, which are explainable/interpretable, whose correct behaviour is guaranteed, and that are robust towards attacks.

The primary research question we consider is to identify the appropriate behavioural models and find whether verification and validation methods exist, or can be developed, for them to scale to medium to large RNN models.  A secondary research question will be how the findings of this analysis can be fed back to improve the RNN efficiently if the analysis uncovers weaknesses.

Guided by a supervisory team with complementary expertise in formal methods as well as deep learning verification and validation, the PhD candidate will have plenty of opportunities to explore and try out new ideas, to write up research outcome in world leading journal and conferences, and to develop transferable skills by attending seminars, workshops and summer schools.

Start Date: 1st October 2023

Further Details:

This PhD project is funded by The Faculty of Science & Engineering at The University of Liverpool and will start on 1st October 2023.

Successful candidates who meet the University of Liverpool eligibility criteria will be awarded a Faculty of Science & Engineering studentship for 3.5 years, covering UK tuition fees and an annual tax-free stipend (e.g. £17,688 p.a. for 2022-23).

Faculty of Science & Engineering students benefit from bespoke graduate training and £5,000 for training, travel and conferences.

The Faculty of Science & Engineering is committed to equality, diversity, widening participation and inclusion. Academic qualifications are considered alongside non-academic experience. Our recruitment process considers potential with the same weighting as past experience. Students must complete a personal statement profoma and ensure this is included in their online application.

How to Apply:

All applicants must complete the personal statement proforma. This is instead of a normal personal/supporting statement/cover letter. The proforma is designed to standardise this part of the application to minimise the difference between those who are given support and those who are not. The proforma can be found here: https://tinyurl.com/ym2ycne4. More information on the application process can be found here: https://tinyurl.com/mwn5952t. When applying online, students should ensure they include the department name in the ‘Programme Applied For’ section of the online form, as well as the Faculty of Science & Engineering as the ‘studentship type’ in the finance section.

Application Web Address: https://www.liverpool.ac.uk/study/postgraduate-research/how-to-apply/