Neural network hardware is coming
I do not understand why AI watchers and pundits do not pay more attention to Elon Musk’s endeavour to build a special purpose neural network computer to assist the self-driving capacities of the newest Teslas. People probably do not realise how new this is: actual neural network hardware has never been build before, and now, within half a year, it will be available in a consumer product.
It is widely known, that graphical processing units from companies like NVIDIA can be used for neural net computations. That is because neural net computations (vector/matrix multiplications) are very much like the computations needed to render images of objects on a screen (it is no coincidence that one of the most successful application areas of neural nets is the classification of images: that is kind of the reverse problem of rendering an object on a screen). But, graphical cards like those made by NVIDIA were never designed only for the purpose of neural net computations. Until now, nobody actually designed and build a special purpose processing unit aimed specifically at this task. And exactly that is what Tesla started with some year ago. The promise is a 1000% performance gain (that means roughly 10 times faster) using the same energy. I am not a believer in the omnipotence of data-driven approaches to AI, but the results of this jump in neural net computation power, coupled with the enormous amounts of data continuously gathered by the roughly half a million Teslas already on the road today, will be very interesting to watch.