Tesla's new patent: Let neural networks automatically adapt to hardware platforms

According to foreign media reports, recently, a Tesla patent titled "System and Method for Adapting a Neural Network Model On a Hardware Platform" was announced.

(Image source: teslarati.com)

Generally speaking, a neural network is a set of algorithms used to collect data and recognize patterns from it. The specific data collected depends on the platform involved and the type of information that the platform can send to the network, such as camera/image data. Differences between platforms mean differences between neural network algorithms, and making adjustments is time-consuming for developers. Just like an application must be programmed to work on the operating system or hardware of a phone or tablet, the same is true for neural networks. However, Tesla takes an automated approach to the adaptability of neural networks.

In the process of adapting a neural network to a specific hardware, software developers must make decisions based on the available options built into the hardware being used. Typically, these options are researched, the hardware documentation reviewed, and the impact analyzed, with each set of options selected and ultimately added to the neural network as a configuration. Tesla calls these options "decision points," and they are a key part of how the invention works.

According to the patent application, after inserting the neural network model and specific hardware platform information to adapt the neural network to the hardware platform, the software code will be spread throughout the neural network to understand the location of the decision points, and then operate the hardware parameters for such decision points to achieve the available configuration. More specifically, the software method looks at the hardware limitations (such as processing resources and performance indicators) and generates settings for the neural network to meet and allow it to run correctly.

The document reads: "In order to implement the abstract neural network, execution decisions may need to be made regarding one or more system components, numerical precision, algorithm selection, data padding, accelerator usage, stride, etc. Such decisions may need to be made on a per-layer or per-tensor basis for the neural network, so a particular neural network may need to make hundreds or even more decisions. The present invention considers many factors before executing the neural network, because the underlying software or hardware platform does not support many configurations, and such configurations may cause the neural network to fail to execute."

Tesla's invention also has the ability to display neural network configuration information on a graphical interface, allowing evaluation and selection in a more user-friendly way. For example, different configurations may require different evaluation times, power consumption, or memory consumption. For example, a configuration can be selected based on the difference between Track Mode and Range Mode, rather than based on how people want the AI ​​to work with the hardware.

The patent application reportedly appears to be one of the products developed after Tesla acquired DeepScale.

DeepScale is an AI startup focused on fully autonomous driving and designing neural networks for small devices. The inventor of the patent is Dr. Michael Driscoll, who was a senior engineer at DeepScale and later became a senior software engineer at Tesla. Before conducting independent research this year, Dr. Forrest Iandola, former CEO of DeepScale, also worked as a senior machine learning scientist at Tesla.