# Leveraging Semantic Embeddings for Safety-Critical Applications

## TL;DR

This paper explores using semantic embeddings to enhance neural network safety in critical applications by enabling interpretability and error detection without modifying existing models.

## Contribution

It introduces a method to generate semantic embeddings from domain knowledge and employs semantic distance as a confidence measure for safety-critical neural network classifiers.

## Key findings

- Semantic distance effectively measures confidence in predictions.
- The approach achieves near state-of-the-art performance.
- It is faster and requires no changes to existing networks.

## Abstract

Semantic Embeddings are a popular way to represent knowledge in the field of zero-shot learning. We observe their interpretability and discuss their potential utility in a safety-critical context. Concretely, we propose to use them to add introspection and error detection capabilities to neural network classifiers. First, we show how to create embeddings from symbolic domain knowledge. We discuss how to use them for interpreting mispredictions and propose a simple error detection scheme. We then introduce the concept of semantic distance: a real-valued score that measures confidence in the semantic space. We evaluate this score on a traffic sign classifier and find that it achieves near state-of-the-art performance, while being significantly faster to compute than other confidence scores. Our approach requires no changes to the original network and is thus applicable to any task for which domain knowledge is available.