http://cs.stanford.edu/people/karpathy/deepimagesent/
I think the impressive thing here is that the GPU is presumably doing GIANT matrix multiplications in real time. A prediction from a neural net is just a series of matrix multiplications, and matrix multiplications are about n^2.8 in complexity, so you can see how matrix multiplications with thousands of rows/columns (often what these sorts of deep image classifiers involve) are hugely computationally expensive.
So it's definitely important for real time machine learning systems to have access to this kind of linear algebra power, but the actual machine learning techniques demonstrated are not super impressive. The hardware is. Which makes sense since this is an Nvidia demo.
However, building a real world working system has challenges that are different to the academic challenge of trying to classify the most classes possible in static images.
[1] http://blogs.technet.com/b/machinelearning/archive/2014/11/1...
If the net learns based on pixels you still have to somehow solve rotation and scale invariance. Or is there something new in deep-learning vs. old-school neural nets that fixes the issues that bedeviled neural nets the first time they were popular?
on the right merc sls classified as SUV
on the left one SUV classified as two VANs
Their algorithm works at about 1Hz rate when doing signs. This is ~state of the art from 20 years ago, but running on small mobile SoC at a slow rate.
