Cross-Modal Scene Networks

TL;DR

This paper introduces methods for learning shared cross-modal scene representations that transfer across different modalities, supported by a new dataset and experiments demonstrating improved transfer and concept alignment.

Contribution

It proposes regularization techniques for cross-modal CNNs to develop modality-agnostic shared representations, a novel dataset, and insights into emergent concept units.

Findings

Shared representations improve cross-modal retrieval

Units activate on consistent concepts regardless of modality

Proposed methods enhance transferability of scene understanding

Abstract

People can recognize scenes across many different modalities beyond natural images. In this paper, we investigate how to learn cross-modal scene representations that transfer across modalities. To study this problem, we introduce a new cross-modal scene dataset. While convolutional neural networks can categorize scenes well, they also learn an intermediate representation not aligned across modalities, which is undesirable for cross-modal transfer applications. We present methods to regularize cross-modal convolutional neural networks so that they have a shared representation that is agnostic of the modality. Our experiments suggest that our scene representation can help transfer representations across modalities for retrieval. Moreover, our visualizations suggest that units emerge in the shared representation that tend to activate on consistent concepts independently of the modality.