Characterizing and overcoming the greedy nature of learning in multi-modal deep neural networks

TL;DR

This paper identifies a greedy learning behavior in multi-modal neural networks where models overly depend on one modality, and proposes a training algorithm to balance learning across modalities, improving generalization.

Contribution

It introduces the concept of conditional utilization rate and a proxy called conditional learning speed to address greedy modality reliance, with an algorithm to balance learning speeds.

Findings

Balanced learning speeds improve model generalization.

Algorithm enhances performance across multiple datasets.

Conditional utilization rate correlates with model dependence on modalities.

Abstract

We hypothesize that due to the greedy nature of learning in multi-modal deep neural networks, these models tend to rely on just one modality while under-fitting the other modalities. Such behavior is counter-intuitive and hurts the models' generalization, as we observe empirically. To estimate the model's dependence on each modality, we compute the gain on the accuracy when the model has access to it in addition to another modality. We refer to this gain as the conditional utilization rate. In the experiments, we consistently observe an imbalance in conditional utilization rates between modalities, across multiple tasks and architectures. Since conditional utilization rate cannot be computed efficiently during training, we introduce a proxy for it based on the pace at which the model learns from each modality, which we refer to as the conditional learning speed. We propose an algorithm to balance the conditional learning speeds between modalities during training and demonstrate that it indeed addresses the issue of greedy learning. The proposed algorithm improves the model's generalization on three datasets: Colored MNIST, ModelNet40, and NVIDIA Dynamic Hand Gesture.