Inspiration from neuroscience -> neural networks.

Convolutional network. Matthew Zeiler & Rob Fergus

Supervised vs unsupervised.

A good principle for learning is for the machine trying to reconstruct the things it wants to learn using its neural net. If what it reconstructs doesn't agree with what it then sees, it should learn. This sounds like learning by imitation.

Regularity helps..

Multimodal, learning combining different kinds of data

Sequence learning and recurrent nets: have memory, can predict sequences (in time say). Can parse words, and they show that grammar can be learned.

Being able to fill gaps in the information you receive (like our brain does, or like machines do with generative models, which also learn) is useful for decision making, as you can know what to expect, even with incomplete info.

Siamese neuronal network Q

Reinforcement learning

Imitation learning

Back-propagation.