# Can Predict/Classify be parallelized?

Being able to parallelize Predict /Classify would be useful on multiple fronts:

• Performance/Scale
• Streaming
• Pedagogical
• Scientific

Performance-wise such parallelization would enable the use of larger datasets and HPC resources; Streaming-wise parallelization would enable certain real-time, distributed, cloud applications; Pedgagogically-wise, parallelization could enhance insights into the structure and judicious choice of predicting/classifying Methods while Scientifically, fine-grained, customized parallelization could promote reproducibility, training variations, extensions of built-in classifiers and the creation of new, hybrid methods.

So are there good frameworks/strategies for parallelizing Predict/Classify?

• Just as an observation, more than just parallelization needs to be addressed to make Predict and Classify good vehicles for machine learning research. The black box function approach, whilst good for a fast start, does not lend itself to tinkering, as far too little of the internals of the algorithms are exposed. – image_doctor Aug 28 '15 at 9:19
• @image_doctor It is possible to peer inside PredictorFunction and ClassifierFunction, which may be of help to both combine results of parallel evaluations of Predict and of Classify and to better understand and perhaps modify their numerical parameters. – bbgodfrey Aug 28 '15 at 13:45

Since Version 11.x a new option is available for Predict, Classify, NetTrain:
TargetDevice -> "GPU"