Skip to content

Learning from small number of examples #2

@sujeongkim

Description

@sujeongkim

I've tried training the network on a single graph. (Training on a single graph and testing on the same graph)
I expected that the network would almost memorize the graph so the loss would be very close to zero. Surprisingly, the loss was about 0.06-0.07, which is more than 20-30 times higher than when I train on 1000 samples). Test loss sometimes is much larger.
I repeated this experiment varying the number of nodes in the graph from 4 to 200, and they all look similar.
I don't have much experience in GNNs and machine learning, but is this normal?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions