| CPC G06N 3/08 (2013.01) [B60W 50/00 (2013.01); G05D 1/0088 (2013.01); B60W 2050/0018 (2013.01)] | 14 Claims |
|
1. A deep network learning apparatus, comprising:
a processor configured to select a deep network model requiring an update in consideration of performance, assign learning amounts for multiple autonomous vehicles in consideration of respective operation patterns of the multiple autonomous vehicles registered through user authentication, distribute the deep network model and learning data to the multiple autonomous vehicles based on the learning amounts for the multiple autonomous vehicles, and receive learning results from the multiple autonomous vehicles; and
a memory configured to store the deep network model and the learning data,
wherein each of the operation patterns includes information about an idle state of an autonomous driving system provided in a corresponding one of the multiple autonomous vehicles,
wherein the processor is configured to differentially assign the learning amounts for the multiple autonomous vehicles considering a period corresponding to the idle state and system resources available in the idle state, and
wherein the processor is configured to calculate reliability evaluation scores by performing verification on the learning results based on validation data, adjust the learning amounts for the multiple autonomous vehicles in consideration of the reliability evaluation scores, and set priorities for the multiple autonomous vehicles from which learning is requested using the reliability evaluation scores.
|
|
8. A deep network learning method performed by a deep network learning apparatus, the method comprising:
selecting a deep network model requiring an update in consideration of performance;
assigning learning amounts for multiple autonomous vehicles in consideration of respective operation patterns of the multiple autonomous vehicles registered through user authentication;
distributing the deep network model and learning data to the multiple autonomous vehicles based on the learning amounts for the multiple autonomous vehicles; and
receiving learning results from the multiple autonomous vehicles,
wherein each of the operation patterns includes information about an idle state of an autonomous driving system provided in a corresponding one of the multiple autonomous vehicles, and
wherein the method further comprises:
differentially assigning the learning amounts for the multiple autonomous vehicles considering a period corresponding to the idle state and system resources available in the idle state,
calculating reliability evaluation scores by performing verification on the learning results based on validation data;
adjusting the learning amounts for the multiple autonomous vehicles in consideration of the reliability evaluation scores; and
setting priorities for the multiple autonomous vehicles from which learning is requested using the reliability evaluation scores.
|