Settings - Multimodal Repr

Model A

Model B

output

paired

Settings - Cross-Modal Distillation

Model A

Model B

Knowledge Distillation

paired

Inference Mode

Settings - Multi-Modal Hashing

Just like the inverse of disentangling.

Settings - Multimodal Repr

Model A

Model B

output

paired

Multimodal Representation

Model A

Model B

Naive method for Joint Representation

Feature A

Feature B

Classifier

Strategy 1 - Early Fusion

Merge

Feature A

Feature B

Classifier B

Strategy 2 - Late Fusion

Ensemble

Classifier A

output

TFN (EMNLP '17)

Feature A

Feature B

Classifier

TFN

Tensor Fusion Network

Use TFN to make feature more complex.

LMF (ACL '18)

Low-rank Multimodal Fusion

LMF = Low Rank Approx. + TFN

Multimodal Deep Learning (ICML '11)

~= cross-modal distillation

= multi-modal repr.

Coordinated Representation

Feature A

Feature B

Classifier

constraint / similarity

• Make Feature A & Feature B more closer.
  • Why works?
    You can think that Feature A & Feature B are distill from each other.
  • What Constraints?
    •  cosine similarity, L2, correlation, etc.

Adaptive Fusion

A GAN method to make two modality be merged

Settings - Cross-Modal Distillation

Model A

Model B

Knowledge Distillation

paired

Inference Mode

Recap: FitNet

• FitNet is two-stages algorithm:
  • Stage 1: pretrain the upper part of student net.
  • Stage 2: Act like original KD (Baseline KD)
• Point: FitNet distill knowledge from features.

Teacher Net (U)

Dataset

Student Net (U)

Regressor

T's feat

S's feat

L2
Loss

Cross modal distillation for supervision transfer (CVPR '16)

Model A

Model B

Knowledge Distillation

paired

Inference Mode

(ImageNet Pretrained)

Modality hallucination (CVPR '16)

幻覺

Model A

Model B

Model B'

Hallucination Network

FitNet
Distillation

Object Detection

Multimodal Knowledge Expansion (CVPR '21)

Model S

Model T

FitNet
Distillation

One of contribution:

Theoretical proved that Model S better than Model T

(pretrained)

Conclusion

• Joint Representation

  • Naive methods like early fusion is widely used nowadays. (Like concatenate, outer product)

• Coordinated Representation

  • One of the branch is just like soft parameter sharing in MTL (Multi-task Learning). (Another is like late-fusion.)

• Cross-Modal Distillation

  • People tries to bring some new KD techniques into cross-modal setting, like BSKD, AT, etc.

• Theoretical Proof?

Comments

• Cross-Modal & Domain Adaptation

  • In fact, different modalities mean different domains, but we have paired data.
  • This implies DA techniques can be applied in cross-modal scenario.

• Cross-Modal & Knowledge Distillation

  • Just like hallucination network, you can adopt various KD techniques and left some degree of freedom. (like IE-KD). 

• Multi-Modal & Multi-task Learning

  • Network Architecture is similar, but reverse way.
  • Joint repr. <-> hard par sharing, and coor repr. <-> soft.

Reference

• Theoretical proof:

  • What Makes Multi-modal Learning Better than Single (Provably) (NIPS '21)

• ​Tutorial:

  • ​Tutorial on Multimodal Machine Learning (ACL '17)

• Implementation for Multimodal Representation
  • Joint Representation
    • Tensor Fusion Network for Multimodal Sentiment Analysis (EMNLP '17)
    • Adaptive Fusion Techniques for Multimodal Data (EACL '21)
  • ​Coordinate Representation
    • ​Deep Canonical Correlation Analysis (ICML '13)
    • ​FitNet + Multi-modal
      • Left some degree of freedom:
        Learning with side information through modality hallucination (CVPR '16) ​
      • Distill from other domain:
        ​Cross modal distillation for supervision transfer (CVPR '16)
    • ​DML + Multi-modal (But I think these kind of paper didn't do DML very well.)
      • ​Towards Cross-Modality Medical Image Segmentation with Online Mutual Knowledge Distillation (AAAI '20)