Projective Latent Interventions for Understanding and Fine-tuning Classifiers

TL;DR

This paper introduces Projective Latent Interventions (PLIs), a method that enables domain experts to intuitively modify and interpret high-dimensional neural network latent spaces by back-propagating manual adjustments in low-dimensional embeddings, improving classifier understanding and performance.

Contribution

The paper proposes PLIs, a novel technique for retraining classifiers through manual modifications of low-dimensional latent embeddings, enhancing interpretability and control in high-dimensional spaces.

Findings

PLIs allow intuitive control over latent decision spaces.

Manual separation of class clusters improves classification performance.

Demonstrated effectiveness on fetal ultrasound imaging data.

Abstract

High-dimensional latent representations learned by neural network classifiers are notoriously hard to interpret. Especially in medical applications, model developers and domain experts desire a better understanding of how these latent representations relate to the resulting classification performance. We present Projective Latent Interventions (PLIs), a technique for retraining classifiers by back-propagating manual changes made to low-dimensional embeddings of the latent space. The back-propagation is based on parametric approximations of t-distributed stochastic neighbourhood embeddings. PLIs allow domain experts to control the latent decision space in an intuitive way in order to better match their expectations. For instance, the performance for specific pairs of classes can be enhanced by manually separating the class clusters in the embedding. We evaluate our technique on a real-world scenario in fetal ultrasound imaging.