A recent patent publication titled 'System and Method for Adapting a Neural Network Model On a Hardware Platform' has provided insight into the way that Tesla aims to create adaptable neural networks that can be used for various hardware platforms.
The patent application is a product of Tesla's acquisition of DeepScale, an AI startup that develops solutions for Full Self Driving as well as neural networks for small devices.
A painstaking review process
Generally speaking, neural network algorithms are trained to carry out a singular task with incredible efficacy. While they allow for patterns to be recognized in data at a rate that humans simply could never achieve, adapting them is often time-consuming for developers.
Typically, when a software developer is adapting a neural network to specific hardware, they have to make decisions based on options built into the hardware. This can result in a painstaking review and configuration process where developers have to review "decision points," as Tesla calls them, in order to make sure the adapted neural network is fit for purpose.
Optimizing neural network adaptability
Tesla's answer to optimizing the adaptability of their neural networks, according to the recent patent filing, is to try to train and automate the sets of algorithms to be adaptable.
Thanks to a system invented by Dr. Michael Driscoll — who was a Senior Staff Engineer for DeepScale before becoming a Senior Software Engineer at Tesla — they might soon be able to do that throughout the company.
According to the patent filing, after plugging a neural network model and specific hardware information, software code analyzes the neural network and pinpoints where the decision points are. The system then runs hardware parameters against that information before providing available configurations.