Features-based embedding or Feature-grounding

TL;DR

This paper explores how to replicate human-like structured reasoning in deep learning by developing feature-grounded embeddings that align with interpretable domain-specific features.

Contribution

It introduces a novel approach to create feature-grounded embeddings that connect shareable representations with interpretable conceptual features.

Findings

Proposes a new method for feature-grounded embedding construction

Aligns deep representations with human-interpretable features

Enhances interpretability of deep learning models

Abstract

In everyday reasoning, when we think about a particular object, we associate it with a unique set of expected properties such as weight, size, or more abstract attributes like density or horsepower. These expectations are shaped by our prior knowledge and the conceptual categories we have formed through experience. This paper investigates how such knowledge-based structured thinking can be reproduced in deep learning models using features based embeddings. Specially, it introduces an specific approach to build feature-grounded embedding, aiming to align shareable representations of operable dictionary with interpretable domain-specific conceptual features.