Exemplary methods and apparatus are provided for configuring a data storage controller to select training data samples from a non-volatile memory (NVM) array for forwarding to an external machine learning processor. The machine learning processor trains a deep neural network model by, e.g., performing various forward and backward passes through a neural network. Within illustrative examples, the data storage controller is equipped with a data sample selection unit that intelligently selects training data stored in the NVM array to forward to the external machine learning processor to reduce an amount of training data to be transferred to the machine learning processor. Among other features, this allows for the practical use of NVM arrays (such as NAND memory arrays) for storing large quantities of machine learning training data, rather than high-speed volatile memory (such as dynamic random access memory), which may be impractical and cost-prohibitive for low-power applications.