ashwhall / meta-learning

Meta-Learning for Few-Shot Class-Incremental Learning

^ That's ridiculous

What do we want (black-box view)?

A system that:

• Can repeatedly have an arbitrary number of new classes added, by only exposing a few images
• Suffers minimal catastrophic interference

Which techniques are a must?

• Meta-learning: Utilise the relationship between train/test splits to learn optimal training methods for few-shot
• Few-shot growth: A minimal number of images required for class-extension

High-Level Approach

1. Train a bunch of models on sub-sets of the training classes
2. Added classes to them (??)
3. Have the meta-learner observe/control steps 1 & 2