The computer is watching
When I was doing an MSc in statistics, back last century, the state of the art in image analysis was recognising handwritten numbers. There was a lot of money in number recognition, for automated sorting of letters by postal code. Australia Post made it easier by having little squares printed on the envelope, showing you where to write the numbers.
In 1997, Terry Pratchett referred to the then state of the art in one of his Discworld novels. A pocket organiser powered by a small demon could do handwriting recognition: you show it a sample, and it says “Yes, that’s handwriting”.
At the time, neural networks weren’t regarded as terribly interesting in statistics. They weren’t good models for the brain, and they were a bit disappointing as black-box classifiers, even discounting how much more black-box they were than the alternatives. The people who taught me were of the opinion that neural networks probably wouldn’t amount to all that much.
It turns out that all they needed was twenty more years development and tens of thousands of times more computing power and training data. Now, neural-network image recognition actually works. I have two posts by Andrej Karpathy to illustrate.
In the first, “What I learned from competing against a ConvNet on ImageNet,” Karpathy tries to do better than a neural network on classifying objects present in photos. He manages. Just. The neural network was particularly good at fine-grained classifications such as breeds of dogs.
The second, “What a Deep Neural Network thinks about your #selfie” indicates some of the problems. The neural network was trained to recognise “good” selfies. Actually, it was trained to recognise selfies that got lots of likes. If you think about what might make a photo get more or fewer likes, you could easily come up with some ideas that aren’t just about photo quality.
Thomas Lumley (@tslumley) is Professor of Biostatistics at the University of Auckland. His research interests include semiparametric models, survey sampling, statistical computing, foundations of statistics, and whatever methodological problems his medical collaborators come up with. He also blogs at Biased and Inefficient See all posts by Thomas Lumley »
An early example of neural networks outperforming humans was in backgammon. Programmers gave up programming the computer player like chess, and instead locked a computer in the room with a backgammon board and a neural network program (not literally) and told it to work it out for itself. Not long after the first big breakthrough in 1992 https://en.wikipedia.org/wiki/TD-Gammon computers were playing better than humans for the first time, and these days the best way to improve your game (as a human) is to play against a neural network in training mode and try to make up heuristics so you can learn “why” it does what it does.
9 years ago