Talking to Intel about AI

• The story that you need GPUs to do machine learning is somewhat of a myth:
  • General purpose CPUs do pretty well if the code is optimised,
  • GPUs limit you in terms of algorithm choice and how you separate your problem, in some cases the limitations are very restrictive,
  • Particularly in general AI, CPUs might be a better platform, unless you want to go all completely neural network.
  • So in general it depends and the devil is in the details (as always ;).
• Libraries/frameworks:
  • TensorFlow.
  • http://deeplearning.net/software/theano/
  • There's also SciKit/Learn.
  • Torch (different approach to TF and Theano, doesn't compile data flow graphs).
  • https://software.intel.com/en-us/ai/deep-learning
  • https://github.com/NervanaSystems/neon - Neural network library.