Learning to Learn via Self-Critique

Antreas Antoniou

Image Classification

• Given:
• Training Data: (x, y)
• Model: f(θ)
• Obtain:
• A trained model f(θ*)
• Target Task:
• A previously unseen test set (or in the wild data)

Current limitations of Image Classification with Deep Learning

• Requires lots labelled data-points.
• Long training times
• Hyper-parameter tuning very expensive
• Architecture design very expensive

ImageNet 2012

Few-Shot Learning

• Learning using only a handful of data-points
• Humans and various other animals routinely exhibit this ability
• Mainstream deep learning performs poorly

​

​

Can you find the defining characteristic of each class?

• Dinosaur vs Ball?
• Dinosaur vs Kids?
• Human Adult vs Human Kids?
• History vs Future?
• Trees vs Cement?
• …
• …
• …

​

So. Which one is it?

What if I asked you to classify the images on the right, given the images on the left?

Support Set

Target Set

Target-set Information

• Based on our little demonstration:
• Target set contains useful task information even though it contains, no label information.
• Support set is important as a reference, but often insufficient.
• Would the target set information be useful to a deep learning model?

Hold that thought.

Few-Shot Learning -

State Of The Art Methods

• Matching Networks/Prototypical Networks
• Meta-Learner LSTM
• SNAIL
• Model Agnostic Meta-Learning variants (MAML/MAML++)
• Latent Embedding Optimization
• Qiao et al 2018
• Baseline ++

Learning only from the Support Set

Why do none of these methods take advantage of the target-set information?

1. Target sets are unlabelled.
2. Existing unsupervised losses unsuitable.
3. Manually inventing such a loss would be a multi-year endeavor.

​

Idea: Let’s automatically meta-learn such an unsupervised loss function overnight on a single computer system.

​

Wait a second..

What is meta-learning?

Enter “Self-Critique and Adapt”

Does the loss network improve the performance of the system?

Thank you!