oneDAL primarily targets algorithms that are extensively used in data analytics. These algorithms typically have many parameters, i.e. knobs to control its internal behavior and produced result. In machine learning, those parameters are often referred as meta-parameters to distinguish them from the model parameters learned during the training. Some algorithms define a dozen meta-parameters, while others depend on another algorithm as, for example, the logistic regression training procedure depends on an optimization algorithm.
Besides meta-parameters, machine learning algorithms may have different stages, such as training and inference. Moreover, the stages of an algorithm may be implemented in a variety of computational methods. For instance, a linear regression model could be trained by solving a system of linear equations [Friedman17] or by applying an iterative optimization solver directly to the empirical risk function [Zhang04].
The same machine learning techniques are often applied for solving problems of different types. In the example with linear regression, the same mathematical model used for solving regression problem is generalized for solving a classification problem, for example, logistic regression. Such techniques differ only in few problem-specific aspects, but share the same subset of meta-parameters and have a common computational flow. oneDAL does not distinguish these techniques into different algorithms. Instead, from oneDAL perspective, the same algorithm may perform different computational tasks.
From computational perspective, algorithm implementation may rely on different
floating-point types, such as
bfloat16. Having a
capability to specify what type is needed is important for the end user as their
precision requirements vary depending on a workload.
To best tackle the mentioned challenges, each algorithm is decomposed into descriptors and operations.