Publications

Heavy-tailed denoising score matching

Published in arXiv, 2021

This paper presents heavy-tailed DSM for score-based models, a generalisation of DSM training which benefits from the use of the generalised normal distribution. This was derived during my PhD with Prof. Pietro Lio at the University of Cambridge and is currently at the pre-print stage.

Recommended citation: Deasy J, Simidjievski N and Lio P (2020). Constraining Variational Inference with Geometric Jensen-Shannon Divergence. arXiv preprint arXiv:2006.10599. http://jacobdeasy.github.io/files/publications/2021-12-17-heavy.pdf

The r-Hunter-Saxton equation, smooth and singular solutions and their approximation

Published in Nonlinearity, 2020

This paper presents the r-Hunter-Saxton equation, a generalisation of the Hunter-Saxton equation derived and analysed during my MSci project with Prof. Colin Cotter at Imperial College London. Currently under review.

Recommended citation: Cotter C, Deasy J, Pryer T (2020). The r-Hunter-Saxton equation, smooth and singular solutions and their approximation. Nonlinearity, Volume 33, Number 12. http://jacobdeasy.github.io/files/publications/2020-10-23-rhs.pdf

Constraining Variational Inference with Geometric Jensen-Shannon Divergence

Published in NeurIPS, 2020

This paper presents geometric Jensen-Shannon VAEs, a generalisation of the beta-VAE family and a closed-form interpolation between forward and reverse Kullback-Liebler divergence. This was derived during my PhD with Prof. Pietro Lio at the University of Cambridge and accepted at NeurIPS 2020.

Adaptive Prediction Timing for Electronic Health Records

Published in ICLR 2020, 2020

This paper presents an adaptive prediction timing method for sequences, based on equi-precision rather than equi-time intervals, applied to electronic health records. This work was carried out during my PhD with Dr Ari Ercole and Prof. Pietro Lio at the University of Cambridge and accepted at the Machine learning in real life workshop at ICLR 2020.

Recommended citation: Deasy J, Ercole A, and Lio P (2020). Adaptive Prediction Timing for Electronic Health Records. arXiv preprint arXiv:2003.02554. http://jacobdeasy.github.io/files/publications/2020-03-05-adaptive.pdf

Time-Sensitive Deep Learning for ICU Outcome Prediction Without Variable Selection or Cleaning

Published in Intensive Care Medicine Experimental, 2019

This paper is about preliminary findings of a time-sensitive and variable-selection-free deep learning model, for mortality prediction in the ICU, developed with Dr Ari Ercole and Prof. Pietro Liò at the University of Cambridge.

Recommended citation: Deasy J, Liò P, Ercole A (2019). Time-Sensitive Deep Learning for ICU Outcome Prediction Without Variable Selection or Cleaning. Intensive Care Medicine Experimental. 2019;1(1):368. http://jacobdeasy.github.io/files/publications/2019-09-27-esicm.pdf

Dynamic survival prediction in intensive care units from heterogeneous time series without the need for variable selection or pre-processing

Published in arXiv, 2019

This paper expands upon prior work on time-sensitive and variable-selection-free deep learning models for mortality prediction in the ICU. Currently under review.

Recommended citation: Deasy J, Liò P, Ercole A (2019). Dynamic survival prediction in intensive care units from heterogeneous time series without the need for variable selection or pre-processing. arXiv preprint arXiv:1909.07214. http://jacobdeasy.github.io/files/publications/2019-09-17-dynamic.pdf

Jacob Deasy