https://www.youtube.com/watch?v=WHNdIuBJHTo&feature=youtu.be
It is able to learn new patterns even if there is a lot of noise. Currently because of the lack of computer power it is only able to learn very simple patterns. With enough computer power I think I could use video to train it.
About the architecture: The pattern recognition engine is a hierarchy of nodes. Each node is responsible for a small field of view, similar to a neuron. Because the input image can be broken up into smaller pieces for processing it is very easy to process it in parallel with multiple computers, which should make it quite fast. Unfortunately I do not have the resources to build a cluster of computers to test the speed so right now this part is just a theory.
I was going to submit this as part of my application to y-combinator but unfortunately I did not complete this prototype on time. I may try to apply for the next one or even do a kickstarter if I can drive enough interest.
Please let me know what you guys think.