https://www.meetup.com/Big-Data-Science/events/rqwdbqybcdbfc/This talk focuses on creation of evolutionary repositories for AI models based on Neural Networks used in deep learning. This presentation describes relational model and RDBMS for storage, retrieval, and for querying with SQL over metadata and data in tables for multiple neural networks. For each pre-trained model, the union over training and test data is saved in normalized tables along with the model. A pre-trained neural network model is generally a well trained neural network model used for a specific function. It is possible to use such a pre-trained model in a composite neural network where one or more other pre-trained models can be also used to form a single rooted directed graph. There is advantage for creating such a composite model to get benefit from other model's intelligence along with saving of time for data preparation and training from scratch. There are definite advantages for creating much simpler pre-trained models first, followed by creations of more complex models by compositions over simpler models. In this talk, we describe relational operations over normalized tables containing training/testing data, metadata along with corresponding pre-trained models in order to derive one or more new composite models. It is possible to define specific set of methods for deriving various composite models with higher complexities in hierarchies, by applying relational operations.
