3 ms·

Actually, I think the Moore's law benefits of continued chip feature shrinkage will be felt most strongly in personal devices. For example, to perform voice rec

Actually, I think the Moore's law benefits of continued chip feature shrinkage will be felt most strongly in personal devices. For example, to perform voice recognition, Siri must send the raw waveform to a cloud server for processing, because the local CPU and RAM are insufficient for acoustic and language modeling. With a couple more CPU generations, i.e. say four years, I expect mobile devices to perform these sort of calculations locally. Beyond speech recognition is machine vision - object/scene detection and recognition. Currently, wireless bandwidth limitations preclude a realtime cloud-based facility for machine vision, but more capable mobile device CPU/GPUs could enable local processing. In general, Moore's law progress enables less expensive cloud servers to perform the same tasks, but that same progress enables mobile devices to perform new categories of tasks!

Someone 15y ago
"Currently, wireless bandwidth limitations preclude a realtime cloud-based facility for machine vision" I do not think that is all what stops it. Just as speech recognition, machine vision is quite different from computer graphics. We _know_ how to make graphics orders of magnitude better; just look at Pixar's latest movie. 'All' we have to do to bring that to your phone is do it more efficiently. For speech recognition and machine vision, I am less optimistic. Reason? I never hear about, say, "almost perfect, but it takes a 1000 CPUs a day to process a second of sound" for speech recognition or object recognition. I wonder why that is? Is it because there are problems we haven't solved yet, or because nobody approaches the problem from that side?
- SlipperySlope 15y ago
  For clarity... I am an AI developer. Recently I added a feature to my dialog system so that if I have not typed anything into the dialog web app in a while, the system begins capturing frames from the webcam and performing face detection using JavaCV. If a face is detected, then face recognition is performed on the clipped greyscale face. If the face best matches mine from the training examples, then the system uses Google speech-to-text synthesis to say hello to me, "hello Stephen". A smartphone cannot practically do this because JavaCV will not run fast enough on an ARM CPU and wireless bandwidth limitations preclude sending whole realtime captured phonecam frames to a server. But, when Moore's Law permits, JavaCV will run on the mobile CPU/GPU combination - and my phone could say hello to me when it recognizes me. I know that Google is working on face-recognition phone-unlocks - but I wonder if their implementation involves a cloud server and one-time sending of a captured frame?

Someone 15y ago

"Currently, wireless bandwidth limitations preclude a realtime cloud-based facility for machine vision" I do not think that is all what stops it. Just as speech recognition, machine vision is quite different from computer graphics. We _know_ how to make graphics orders of magnitude better; just look at Pixar's latest movie. 'All' we have to do to bring that to your phone is do it more efficiently. For speech recognition and machine vision, I am less optimistic. Reason? I never hear about, say, "almost perfect, but it takes a 1000 CPUs a day to process a second of sound" for speech recognition or object recognition. I wonder why that is? Is it because there are problems we haven't solved yet, or because nobody approaches the problem from that side?

SlipperySlope 15y ago
For clarity... I am an AI developer. Recently I added a feature to my dialog system so that if I have not typed anything into the dialog web app in a while, the system begins capturing frames from the webcam and performing face detection using JavaCV. If a face is detected, then face recognition is performed on the clipped greyscale face. If the face best matches mine from the training examples, then the system uses Google speech-to-text synthesis to say hello to me, "hello Stephen". A smartphone cannot practically do this because JavaCV will not run fast enough on an ARM CPU and wireless bandwidth limitations preclude sending whole realtime captured phonecam frames to a server. But, when Moore's Law permits, JavaCV will run on the mobile CPU/GPU combination - and my phone could say hello to me when it recognizes me. I know that Google is working on face-recognition phone-unlocks - but I wonder if their implementation involves a cloud server and one-time sending of a captured frame?