Machine learning — Hierarchical loss for neural network classifiers

Hypothesis

Training a classifier directly on a hierarchical loss, which penalises confusing a dog with a skyscraper more than confusing two dog breeds, would reduce serious mistakes more than standard training.

What failed

In extensive experiments, standard cross-entropy training reduced the hierarchical loss almost as much as training on the hierarchical loss directly.

Broken assumption

Optimising a hierarchy-aware loss directly is a reliable way to make a classifier's mistakes less severe.

Why

Plain stochastic gradient descent from random starting points already lowers the hierarchical loss as a side effect, so the new objective gives little extra benefit.

From the published paper: Wu et al. PLoS ONE, 2019. ↗CC BY