Re:Machine Learning with TensorFlow Enabled Mobile Proof-Of-...
by candiodari on Tuesday October 17, 2017 @07:11AM
(#15490029)
Math heavy ML is simple: ∀ problems: solution is Convnets (CNN in the article).
But yes, you should learn to apply multilayer convnets to image problems.
And don't worry about architecture. There's 2 variables : number of parameters and number of layers. There's an optimum, which will be bigger than "advanced" architectures, but it will match the advanced architectures' performance very closely (1-2% worse perhaps).
Slight issue is that it takes 2-3 hours to test a (number of free parameters, number of layers) combination, and so "hyperparameter tuning" (pick random and test, repeat) takes a long time. But if you make it a nice pipeline you can do it on cloud in parallel on a regular basis (as new data gets added) and have extremely good results.
But yes, you should learn to apply multilayer convnets to image problems.
And don't worry about architecture. There's 2 variables : number of parameters and number of layers. There's an optimum, which will be bigger than "advanced" architectures, but it will match the advanced architectures' performance very closely (1-2% worse perhaps).
Slight issue is that it takes 2-3 hours to test a (number of free parameters, number of layers) combination, and so "hyperparameter tuning" (pick random and test, repeat) takes a long time. But if you make it a nice pipeline you can do it on cloud in parallel on a regular basis (as new data gets added) and have extremely good results.