Computers now better than humans at recognising and sorting images
|
Version 0 of 1. First, the robots beat us at assembling cars. Then, they beat us at playing chess. Now, they’ve also got better than us at sorting images into predefined categories. It might not sound like much, but the success of the Minwa supercomputer, which can sort a million images into a thousand predefined categories with an error rate less than the typical human, makes it the latest secret weapon of the company known as “China’s Google”, Baidu. Minwa is able to scan the massive ImageNet picture database and sort the images into the correct category – telling the difference between different breeds of dog, for instance – with an error rate of just 4.58%. For comparison, the typical human gets around 5%, while software from Microsoft and Google has already reached 4.94% and 4.8% respectively. In a paper presenting its results, Baidu details just how its AI achieved the new record. The software runs on a massive cluster of supercomputers, using 36 server nodes which between them have 72 processors with 432 cores, as well as 144 GPUs (graphics processing units, high-performance specialised chips typically used to deal with visual data). It applies a “neural network” to recognise the images, training the software up with high-resolution versions of pictures so that it develops an understanding of the properties that it is looking for. To prevent “over fitting”, when Minwa’s mathematical model becomes too obsessed with the specifics of the portion of images it is trained on, the data is also supplied in skewed forms, with vignetting, cropping, colour and shape distortion all applied to make sure the model only learns the important characteristics of the image subject. Baidu As a result, the model is so robust that it can even recognise the subject of an image when it is printed out, held at an oblique angle, and then photographed a second time. Andrew Ng, Baidu’s chief scientist, told the Wall Street Journal: “I am very excited about all the progress in computer vision that the whole community has made. Computers can understand images so much better and do so many things that they couldn’t do just a year ago.” The company is also using an even larger supercomputer cluster to analyse speech data to improve its voice recognition. If you want to try your own hand at recognising images, an interactive version of the same test that Minwa gets right 95.42% of the time is available online. |