Computer's imagination creates human faces

摘要

Ever wondered what computers think we look like? An artificial neural network has been trained to conjure realistic-looking images on demand, including human faces. Artificial neural networks are used to pick out specific characteristics from large datasets, such as simple geometric shapes, or more complex figures like cats or faces. They learn by analysing training data like photos, but exactly how different components of a network learn to recognise particular objects is rather opaque. Google's psychedelic DeepDream landscapes are a recent example: in these, the computer "sees" dogs, birds and other objects that aren't really there.