to benchmark learning techniques. It is followed by a short discussion on exchanging model quality and model footprint. An in-depth discussion of data augmentation and distillation follows right after. discuss a few label invariant image and text transformation techniques. Image Transformations This discussion is organized into the following two categories: spatial transformation and value transformation cheaper than human labor costs to produce training samples. We have chosen four techniques for deeper discussion to cover both statistical and deep learning based synthetic data generation models. Back Translation

Ridgeway G . Special Invited Paper. Additive Logistic Regression: A Statistical View of Boosting: Discussion[J]. Annals of Statistics, 2000, 28(2):393-400. [6] FRIEDMAN J H . Stochastic gradient boosting[J]

i=1 1(y(i) = y) (23) Remark: We assume binary features (Xj ∈ {0, 1} for ∀j ∈ [n]) in the above discussion. What if Xj ∈ {1, 2, · · · , v}? Can we get similar results? Check it by yourselves! 4.4 Laplace

readers to play with other choices and see how they affect the results. It's now time to conclude our discussion on automation with a short introduction to Automated ML or AutoML in the final section. Summary

an n-dimensional matrix) to denote inputs, weights and the bias respectively. To simplify this discussion, let’s assume the shape (an array describing the size of each dimension) of X as [batch size, D1]