Bioinformatics Seminar
Time: 11AM
Venue: Davis Auditorium and Teams
10 May 2022
Interpretable deep metric learning models for small datasets
Jarryd MartinWEHI Bioinformatics
Many datasets of interest in Bioinformatics have two things in common: they have many covariates and very few samples. Interpretable, accurate models for these datasets are sorely needed. Unfortunately, machine learning methods typically need thousands of samples to perform adequately, and this severely limits their applicability. Deep neural network models have one desirable property in this context: they can pretrain on other (larger) datasets, before finetuning on a small dataset of interest. However, typical neural network models are not interpretable. In this talk I will describe an approach to building two kinds of interpretability into neural network models. Models with ‘feature interpretability’ provide explanations for a sample prediction in terms of that sample’s input covariates. Models that offer ‘case-based interpretability’ provide explanations for a sample prediction in terms of similarity to other samples in the dataset. I will present a deep metric learning model, Attentive Neighbourhood Components Analysis, that incorporates both these forms of interpretability, and a novel extension of this approach to survival analysis datasets.