Chapter 11: Streamlining Network Implementation with AutoML
Computer vision, particularly when combined with deep learning, is a field that's not suitable for the faint of heart! While in traditional computer programming, we have a limited set of options for debugging and experimentation, this is not the case in machine learning.
Of course, the stochastic nature of machine learning itself plays a role in making the process of creating a good enough solution difficult, but so do the myriad of parameters, variables, knobs, and settings we need to get right to unlock the true power of a neural network for a particular problem.
Selecting a proper architecture is just the beginning because we also need to consider preprocessing techniques, learning rates, optimizers, loss functions, and data splits, among a multiplicity of other factors.
My point is that deep learning is hard! Where do you start? Wouldn't it be great if we had a way to ease the burden of searching through...
