Daily Dose of Paper : Overcoming Catastrophic Forgetting in Graph Neural Networks

 

Overcoming Catastrophic Forgetting in Graph Neural Networks

AAAI 2021 paper.

TLDR

wsnalgo
Designed continual learning strategy tailored for GNN,
where most previous works have been about CNN-based networks.
Penalize updating important parameter in terms of loss change and attention change.

Quick Look

Research Topic

  • Category (General): Graph Neural Network, Continual Learning
  • Category (Specific):

Paper Summary (What)

  • Problem statement
    • Previous continual learning approaches were not for GNN-based models.
  • Solutions
    • Define two importance metrics, $I_{loss}$ and $I_{attention}$.
    • $I_{loss}$ : $\sum_{m} \frac{\partial L}{\partial w_m} \Delta w_m$
    • $I_{attention}$ : Given graph attention between node $i$ and $j$ as $e^{(l)}{ij} = a(H^{(l-1)}{i,j} ; W^{(l)})$,
      $e^{(l)}{i} = [e^{(l)}{i1}, \cdots , e^{(l)}{i N_i}]$,
      $g_m(H^{(l-1)}) = \frac{\partial ([e^{(l)}
      {1}, \cdots , e^{(l)}{D}])}{\partial w_m}$,
      $I
      {attention} = [|| g_m(H^{(l-1)}) ||]$
    • In short, $I_{attention}$ is the L2 norm of change of attention.
  1. Penalize updating important parameters with loss term.

eq1

Notable References

Thoughts

As it is said in the paper, no objective function for attention change.
Thus, it is just penalizing the attention change as a whole,
which is not ‘topology preserving’ way to freeze,
but rather a way to freeze all graph structure.
Is it really a desirable way of freezing? May be more clever topological aware design?

Footnote

아래와 같은 양식을 활용한다.

# 
## Quick Look
- **Authors & Affiliation**: [Authors][Affiliations]
- **Link**: [Paper link]
- **Comments**: [e.g. Published at X / arXiv paper / in review.]
- **TLDR**: [One or at most two line summary]
- **Relevance**: [Score between 1 and 5, stating how relevant this paper is to your work. Usually filled in at the end.]
# Research Topic
- **Category** (General):
- **Category** (Specific):
# Paper Summary (What)
[Summary of the paper - a few sentences with bullet points. What did they do?]

wbjeon2k

Pursuite for Progress.

This work is licensed under a Attribution-NonCommercial 4.0 International license. Attribution-NonCommercial 4.0 International