The Modality Focusing Hypothesis: On the Blink of Multimodal Knowledge Distillation

06/13/2022
by   Zihui Xue, et al.
5

Multimodal knowledge distillation (KD) extends traditional knowledge distillation to the area of multimodal learning. One common practice is to adopt a well-performed multimodal network as the teacher in the hope that it can transfer its full knowledge to a unimodal student for performance improvement. In this paper, we investigate the efficacy of multimodal KD. We begin by providing two failure cases of it and demonstrate that KD is not a universal cure in multimodal knowledge transfer. We present the modality Venn diagram to understand modality relationships and the modality focusing hypothesis revealing the decisive factor in the efficacy of multimodal KD. Experimental results on 6 multimodal datasets help justify our hypothesis, diagnose failure cases, and point directions to improve distillation performance.

READ FULL TEXT
research
03/26/2021

Multimodal Knowledge Expansion

The popularity of multimodal sensors and the accessibility of the Intern...
research
10/15/2021

From Multimodal to Unimodal Attention in Transformers using Knowledge Distillation

Multimodal Deep Learning has garnered much interest, and transformers ha...
research
03/24/2023

Decoupled Multimodal Distilling for Emotion Recognition

Human multimodal emotion recognition (MER) aims to perceive human emotio...
research
08/17/2022

Leukocyte Classification using Multimodal Architecture Enhanced by Knowledge Distillation

Recently, a lot of automated white blood cells (WBC) or leukocyte classi...
research
01/06/2021

Modality-specific Distillation

Large neural networks are impractical to deploy on mobile devices due to...
research
07/14/2023

Multimodal Distillation for Egocentric Action Recognition

The focal point of egocentric video understanding is modelling hand-obje...
research
09/06/2018

RDPD: Rich Data Helps Poor Data via Imitation

In many situations, we have both rich- and poor- data environments: in a...

Please sign up or login with your details

Forgot password? Click here to reset