Driving Safe with Autopilot: Tesla Introduced Patent for Neural Network System with Hardware Adaptability

The target of this article is to provide readers with info on the steps Tesla takes to design adaptable neural networks that can be implemented for various hardware platforms. Readers will go deeper into the following topics:

· The use of neural network algorithms for autopilot;

· The model configuration method and its steps;

· The reasons manufacturers need such a patent.

So, in case all those things are of great interest to you, keep reading the article. We’d like to mention that if you want to develop a thing like Tesla for your personal startup, find automotive software engineers at Engre platform . Engineering talents from Engre truly create magic!

About Patent Developed by Tesla Team

Recently, Tesla revealed the patent publication called “'System and Method for Adapting a Neural Network Model on a Hardware Platform.” What is it about? The team wanted to give insight into how they plan to design adaptable neural networks that may become an integral part of a variety of hardware platforms.

This patent was successfully prepared as a part of the DeepScale solution . The last is considered a powerful AI startup whose mission is to design applications for Autopilot as well as Full Self-Driving capabilities including neural networks for various devices of small sizes.

For those who are not on the topic, let us clarify the following.

In general, neural network algorithms are trained to execute a certain task with a super high level of efficiency. Such algorithms allow getting patterns recognized in data at a speed that engineering professionals just could never gain. However, the point is it is a truly time-consuming procedure for engineers to adapt to those algorithms.

Traditionally, when a software engineer tries to adapt a neural network to certain hardware, their work strongly depends on the configurations inside the hardware. That, as a rule, leads to a slow, painstaking, and complicated procedure. This means automotive engineers are to analyze the so-called decision points to guarantee that the adapted neural network ideally fits the initial goal.

By the way, that is the main thing for consideration by engineers when it comes to safety measures in electric vehicle development. See the list of the latest achievements in the industry including new electric vehicles by General Motors to review the level manufacturers have already achieved in designing such transport.

Optimizing Neural Network Algorithms

Here, the question raises how to optimize the neural network algorithms adaptability. According to patent, Tesla’s answer sounds like they should train as well as automate the specific algorithm sets. It is possible due to the solution designed by Dr. Michael Driscoll, Senior Software Engineer at Tesla who also worked on the DeepScale startup.

According to Tesla, various neural networks are usually trained with several hyperparameters. Then, the last is implemented to examine the same training set for validation. They choose a certain neural network to be used in the future. The choice depends on the required performance and the sensibility parameters of specific applications.

For applications based on Machine Learning, they have to create neural networks on platforms prepared earlier. Note that developing a neural network for a certain application is usually a hard process. As it was previously mentioned, different neural networks possess different requirements (hardware/software constituents).This places challenging restrictions on configurations.

The problem is challenging because engineers need to figure out which data layout type to implement, which algorithms to use, and so on. That depends on all the options available for each configuration kind. This leads to the earlier mentioned issue of decision points because a decision at any certain decision point may lead to the state when the configuration of models is ineffective.

According to the patent by Tesla, they stress on the thing that the realization of model configuration methods, as well as systems, involves techniques for determining neural network configurations adapted to a certain platform. Let’s now consider what model configuration system and model configuration method by Tesla means to understand how they are interrelated.

The model configuration system includes the following:

· a neural network model;

· a model configuration platform;

· constraints as well as traversal module;

· a constraint satisfaction solver;

· a configuration as well as performance module;

· a data store;

· a hardware platform.

The model configuration method contains five steps:

· Step 1: to identify decision points they should go through a neural network model;

· Step 2: working with model configuration system that involves identifying model, performance, and variable constraints;

· Step 3: designing SMT (Satisfiability Modulo Theories) solver for the neural network model;

· Step 4: clarifying that the chosen configurations ideally fit the purpose;

· Step 5: finding out the configuration that fits all the required performance metrics.

In the patent publication, they point out that the combination of the system and the method may use any combination and transformation of different system constituents and method procedures. This supports engineers while they clarify the effectively chosen configurations. They implement a constraint satisfaction solver (SAT or SMT solver) and optionally decide on configurations that perfectly fit one or more required performance criteria of neural network models.

Tesla explains that there are a certain number of tasks that the neural network model of the system executes depending on the combinations.

For example, in some combinations, the neural network model is a model that is set on the same PC as the model configuration platform. Meanwhile, there are combinations where the neural network model, model configuration platform, and hardware platform belong to separate devices. Additionally, in some combinations, the neural network model performs commands based on machine learning or deep learning approaches.

Tesla's system will provide automotive engineers with noticing the available configurations on an easy-to-use display. This simplifies viewing and selecting the right direction to proceed.