Tesla’s self-driving patent application hints at AI safety improvements
A recently published Tesla patent application titled “System and Method for Handling Errors in a Vehicle Neural Network Processor” describes a way to safely handle errors encountered in self-driving software. Rather than risking delays in driving responses that result from input data errors, a signal is sent to ignore the bad information and continue processing as usual. Tesla’s application was published May 23, 2019 as International Publication No. WO/2019/099941.
Tesla’s patent application describes the issue as follows:
“Some types of errors may cause neural network processor to hang or time out. That is, one or more portions of neural network processor may freeze or otherwise remain inactive for more than a predetermined amount of time. When a timeout error is encountered, [the] neural network processor may cease to provide output data and/or respond to input data. Other types of errors, such as program errors and/or data errors, may cause the output data generated by [the] neural network processor to be corrupted. When such errors are encountered, [the] neural network processor may continue to provide output data, but the result may be incorrect, meaningless, and/or otherwise unusable.”
Tesla's self-driving patent application hints at AI safety improvements
A recently published Tesla patent application titled “System and Method for Handling Errors in a Vehicle Neural Network Processor” describes a way to safely handle errors encountered in self-driving software. Rather than risking delays in driving responses that result from input data errors, a signal is sent to ignore the bad information and continue processing as usual. Tesla’s application was published May 23, 2019 as International Publication No. WO/2019/099941.
Tesla’s patent application describes the issue as follows:
“Some types of errors may cause neural network processor to hang or time out. That is, one or more portions of neural network processor may freeze or otherwise remain inactive for more than a predetermined amount of time. When a timeout error is encountered, [the] neural network processor may cease to provide output data and/or respond to input data. Other types of errors, such as program errors and/or data errors, may cause the output data generated by [the] neural network processor to be corrupted. When such errors are encountered, [the] neural network processor may continue to provide output data, but the result may be incorrect, meaningless, and/or otherwise unusable.”
Tesla's self-driving patent application hints at AI safety improvements
