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

Please sign up or login with your details

Continue with:
Or login with email
Enter Password
Re-enter Password

Forgot password? Click here to reset
Success!
Error Icon An error occurred

Sign in with Google

×

Use your Google Account to sign in to DeepAI

×
Pro

Consider DeepAI Pro

Subscribe to DeepAI Pro
DeepAI Pro
Provides a limited generation allowance each month. When exceeded, you are charged overage rates available at deepai.org/pricing. Also includes an ad-free experience and API access. Renews automatically until canceled. Non-refundable.
Subtotal
Total due today

Payment

Add DeepAI credits
DeepAI credits
One-time purchase. Credits are added to your wallet after payment.
Subtotal
Total due today

Payment