ACM TechNews linked to an article, "How Many Computers to Identify a Cat? 16,000."
It turns out that our professor is doing some work on the side, leading a research project with Google X laboratory. The article explains that they built a neural network. Unsupervised, it learned on its own the concept of a cat, along with 20,000 other concepts. An interesting comment was, "The Google brain assembled a dreamlike digital image of a cat by
employing a hierarchy of memory locations to successively cull out
general features after being exposed to millions of images." I'm pretty sure Ng didn't explain how to cull out general features in the neural network lectures.