TOP: A Compiler-Based Framework for Optimizing Machine Learning Algorithms through Generalized Triangle Inequality

This paper describes our recent research progress on generalizing triangle inequality (TI) to optimize Machine Learning algorithms that involve either vector dot products (e.g., Neural Networks) or distance calculations (e.g., KNN, KMeans). The progress includes a new form of TI named Angular Triangular Inequality, abstractions to enable unified treatment to various ML algorithms, and TOP, a compilerbased optimizer for effectively applying TI to optimize machine learning algorithms. Experiments show that TOP is able to automatically produce optimized algorithms that either matches or outperforms manually designed algorithms, giving up to 237x speedups and 2.5X on average.