Bio
Torty began her Doctorate at the Turing in September 2019. She was based at the University of Bristol where she also completed her MEng in Computer Science in June 2018. Throughout her degree, Torty became interested in the negative impacts of technology on society, culminating in her master’s thesis, “Towards ethical moral machines”. The thesis, grounded in the autonomous vehicle domain, assesses recent attempts to model human morality and particularly scrutinises the effects of model assumptions on results.
Research interests
Torty’s PhD research is sumarised as follows:
Given an AI system, which perhaps predicts the probability of lung cancer from several lifestyle factors, a post-hoc local explanation is one of the form `this patient was classified as at risk of lung cancer because their bloodl in terms of their input features. Post-hoc local explanations have been widely successful for explaining image and tabular data yet explaining time series data has been relatively under-ex pressure is in the 95th percentile', i.e. it explains an AI system's output for an individuaplored. Torty's research bridges the interpretability gap between existing ways of generating post-hoc local explanations for image and tabular data to time series data. They bring together game theoretic, causal and statistical ideas to develop four methods for explaining time series, uniting counterfactuals, differential attribution, functional decomposition and frequency domain analysis.
Selected publications and papers
'- Sivill, Torty, and Peter Flach. "LIMESegment: Meaningful, Realistic Time Series Explanations." International Conference on Artificial Intelligence and Statistics. PMLR, 2022.
- Torty Sivill, Vanja Ljevar, James Goulding, Anya Skatova. "What Can Transactional Data Reveal About the Prevalence of Menstrual Pain in England?" Digital Footprints Conference 2023
- Delivered presentation `Explaining Explainable AI' at the Jean Golding Institute Data Showcase 2022, LV Insurance, University of Bristol Faculty of Law