Recurrent Neural Networks

This Self Study was planned as a learning tool to understand the implementation of neural networks using Nvidia AI Toolkit, (now part of the Nvidia cuDNN Deep Neural Network API). The goal was to develop a proposal for use of this tooling by a leading insurance organization with an actuary department skilled at the use of Excel and Mote Carlo simulation for product actuary. The result was an Excel wrapper add-in for Excel which ran on CUDA-equipped devices (GPUs) accessed via the toolkit and a C++ CLI wrapper. The characteristics of the platform was evaluated against open quote engines in multiple fields. The use of a recurrent network was shown to be valid for storage historical domain vectors for evaluation of new inquiries. Preliminary results indicated a 14% lower average product quote with ~89% test accuracy, deliverable for authorization in less than a week, where prior methods took many months of testing and approval to produce.

Modern "NPU" architectures such as Copilot+PC and other major hardware venders are a successor to the groundwork laid down by the NVidia CUDA Toolkit many years ago, paving the way for highly personalized AI assistants in the near future.