Hugo Larochelle - Meta Learning

Humans seem to be able to transfer their knowledge (use prior knowledge to learn new tasks) much better than machine learning algorithms for now.

Human-level concept learning through probabilistic program induction

Meta learning references

The train and validation set are (each) a set of learning tasks (for instance classification on different classes (different training classes, different test classes in each task)).

Prototypical Networks for Few-shot Learning

Meta Learner LSTM

Learns how to update the parameters of the model and their initialization using an LSTM

MAML <-- seems to be a must-read

• Removed the LSTM and do just one-step gradient update
• Showed that just learning a good initialization is enough.
• Showed improvement in decoupling bias and weights meta-learning

A Simple Neural Attentive Meta-Learner