Skip to content

Run an evolutionary training run using evotorch #2

Description

@mekaneeky

My hypothesis is since the space of loss values from different weights is less informative in low-resource domains due to less data. Getting directions/gradients in the loss space might be less clear, similar to deep RL. While random search would be too inefficient. Evolutionary search over the space of parameters would still be guided by the training loss, but would have less directionality, therefore might stumble on a better solution for low data domains.

Possibly a long shot, but not too complex.

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