Neural networks: AI’S “animal brain”
At the core of the deepfakes code is a “deep neural network” – a computing system vaguely inspired by the biological neural networks that make up animal brains. Such systems “learn” – or progressively improve their performance – by analysing vast amounts of data, acquainting themselves with the information via trial and error, and developing something like human flexibility; rather than needing to be preprogrammed with fixed rules, they rewire themselves by absorbing patterns in the data. Neural networks have driven many striking recent improvements in artificial intelligence, in areas such as translation, speech recognition and image recognition. Fakeapp uses a suite of neural networking tools that were developed by Google’s AI division and released to the public in 2015. The software teaches itself to perform image recognition tasks through trial and error. First, Fakeapp trains itself, using “training data” in the form of photos and videos. Then it stitches the face onto another head on a video clip – accurately preserving the facial expression on the original video. These technologies have been developed by online communities, where developers are often happy to share techniques; the pace of progress is fast.