Implementation Roadmap for Neural Networks in Array Databases

Relational databases lack behind when handling array data and thus array databases were created to fill this gap. Array databases provide optimized storage, retrieval, and processing of multidimensional discrete data (MDD), also known as array data. Just like relational array databases, data processing in array databases is handled declaratively through an array query language that offers enough expressible power to define a myriad of operations. However, despite the advancements in array database technology, there is still a gap in describing machine learning (ML) algorithms and in particular neural networks which, in recent years, have been adopted for predicting phenomena in science and engineering. In this contribution, we outline an implementation roadmap for defining neural networks in an array database. We first identify the necessary linear algebra operators present in a feed-forward neural network and use them to define the training and prediction operations of that network. We also define other operators that, though they are not part of linear algebra, are essential for a complete machine-learning implementation.