Learning Aligned Cross-Modal Representations from Weakly Aligned Data

TLDR

Scene recognition works across modalities, yet CNNs learn representations that are not aligned across modalities, limiting cross‑modal transfer. The study investigates how to learn cross‑modal scene representations that transfer across modalities. We introduce a new cross‑modal scene dataset and regularize CNNs to learn a modality‑agnostic shared representation. Experiments show the shared representation improves cross‑modal retrieval, and visualizations reveal units that activate on consistent concepts regardless of modality.

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 cross-modal scenes well, they also learn an intermediate representation not aligned across modalities, which is undesirable for crossmodal 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.