Back to results
US 12,743,632 B2Granted
Debugging in federated learning systems
AssigneeCisco Technology, Inc.
InventorsJayanth Srinivasa, Myungjin Lee, Ramana Rao V. R. Kompella
Life of this patent
Estimating the term…
Abstract
In one embodiment, a device makes a determination that performance of a global model generated by aggregating local models trained by a plurality of trainer nodes in a federated learning system has experienced a degradation. The device selects, in response to the determination, a particular trainer node from among the plurality of trainer nodes to generate debugging metrics. The device provides an indication that the particular trainer node is a root cause of the degradation.
Claims20
A method, comprising:
- making, by a device, a determination that performance of a global model generated by a federated learning system has experienced a degradation, wherein the global model was generated by aggregating local models trained by a plurality of trainer nodes in the federated learning system;
- obtaining first validation data related to validation of the global model:
- obtaining second validation data related to validation of a local model associated with a particular trainer node from among the plurality of trainer nodes;
- comparing the first validation data to the second validation data to determine a degree of influence the local model associated with the particular trainer node had on the global model;
- selecting, by the device, the particular trainer node from among the plurality of trainer nodes based, at least in part, on the degree of influence being greater than a value;
- obtaining, by the device, debugging metrics from the particular trainer node based, at least in part, on selecting the particular trainer node;
- and
- providing, by the device, an indication that the particular trainer node is a root cause of the degradation.
5 more elements
An apparatus, comprising:
- one or more network interfaces;
- a processor coupled to the one or more network interfaces;
- and
- a memory configured to store instructions that, when executed by the processor, configure the processor to:
- make a determination that performance of a global model generated by a federated learning system has experienced a degradation, wherein the global model was generated by aggregating local models trained by a plurality of trainer nodes in the federated learning system;
- obtain first validation data related to validation of the global model;
- obtain second validation data related to validation of a local model associated with a particular trainer node from among the plurality of trainer nodes;
- compare the first validation data to the second validation data to determine a degree of influence the local model associated with the particular trainer node had on the global model;
- select the particular trainer node from among the plurality of trainer nodes based, at least in part, on the degree of influence being greater than a value;
- obtain debugging metrics from the particular trainer node;
- and
- provide an indication that the particular trainer node is a root cause of the degradation.
9 more elements
A tangible, non-transitory, computer-readable medium storing program instructions that cause a device of a federated learning system to execute a process comprising:
- making, by the device, a determination that performance of a global model generated by the federated learning system has experienced a degradation, wherein the global model was generated by aggregating local models trained by a plurality of trainer nodes in the federated learning system;
- obtaining first validation data related to validation of the global model;
- obtaining second validation data related to validation of a local model associated with a first trainer node from among the plurality of trainer nodes;
- analyzing the first validation data and the second validation data;
- determining a degree of influence the local model associated with the first trainer node had on the global model;
- selecting, by the device, the first trainer node from among the plurality of trainer nodes based, at least in part, on the degree of influence satisfying one or more criteria;
- obtaining, by the device, one or more debugging metrics from the first trainer node;
- and
- providing, by the device, an indication that the first trainer node is a root cause of the degradation.
6 more elements
Dependencies are inferred from the claim text.