Monitoring machine learning models in resource constrained settings
A system subscribes to one or more inference topics to which inferences are published on behalf of trainable software models executing in a vehicle computing environment. The system receives inferences from the topics as the inferences are published to the topics and associates the inferences with one or more trainable software models to be monitored. Also, the system identifies instances of unexpected output based on comparison of received inferences, associated with a given model to be monitored, to expected inference values identified in a configuration file, stored in a vehicle memory and associated with the given model, and, responsive to identifying the unexpected output, devises a modification strategy for the model based on characteristics of the unexpected output.
1 . A system comprising:
one or more processors configured to:
subscribe to one or more inference topics to which inferences are published on behalf of trainable software models executing in a vehicle computing environment;
receive inferences from the topics as the inferences are published to the topics;
associate the inferences with one or more trainable software models to be monitored;
identify instances of unexpected output based on comparison of received inferences, associated with a given trainable software model to be monitored, to expected inference values identified in a configuration file, the expected inference values based at least in part on values provided from a remote server to the vehicle indicating historical output values previously generated by variants of the given trainable software model executing in other vehicles, stored in a vehicle memory, and associated with the given trainable software model; and
responsive to identifying the unexpected output, devise a modification strategy for the given trainable software model based on characteristics of the unexpected output.
2 . The system of claim 1 , wherein the inference topics are identified based on information received responsive to a configuration file, associated with the given trainable software model to be monitored, requesting loading of the given trainable software model for execution.
3 . The system of claim 1 , wherein at least one of the one or more processors is configured to associate the inferences based on a configuration file associated with each given trainable software model to be monitored, indicating applicable inferences.
4 . The system of claim 1 , wherein the modification strategy includes training the given trainable software model including varying weighting of inputs to the given trainable software model based on interpretability data derived from the inferences associated with the given trainable software model, leading to the unexpected output and identifying inputs to the given trainable software model that skew the inference to produce the unexpected results.
5 . The system of claim 1 , wherein the modification strategy includes reversion to a prior version of the given trainable software model, stored in vehicle memory, providing output closer to expectations determined based on the unexpected output compared to prior output of the prior version of the given trainable software model, observed and stored in vehicle memory.
6 . The system of claim 1 , wherein at least one of the one or more processors is further configured to provide interpretability reporting, based on identification of unexpectedly impactful input values of inputs identified by the configuration file for the given trainable software model as having above a defined threshold impact on the inferences resulting from the given trainable software model.
7 . The system of claim 1 , wherein the expected inference values used in the identification of unexpected output are further based on vehicle-stored historical output values previously generated by the given trainable software model or a prior version of the given trainable software model.
8 . A method comprising:
subscribing to one or more inference topics to which inferences are published on behalf of trainable software models executing in a vehicle computing environment;
receiving inferences from the topics as the inferences are published to the topics;
associating the inferences with one or more trainable software models to be monitored;
identifying instances of unexpected output based on comparison of received inferences, associated with a given trainable software model to be monitored, to expected inference values identified in a configuration file, the expected inference values based at least in part on values provided from a remote server to the vehicle indicating historical output values previously generated by variants of the given trainable software model executing in other vehicles, stored in a vehicle memory, and associated with the given trainable software model; and
responsive to identifying the unexpected output, devising a modification, strategy for the given trainable software model based on characteristics of the unexpected output.
9 . The method of claim 8 , wherein the inference topics are identified based on information received responsive to a configuration file, associated with the given trainable software model to be monitored, requesting, loading of the given trainable software model for execution.
10 . The method of claim 8 , wherein the associating further comprises associating the inferences based on a configuration file associated with each given trainable software model to be monitored, indicating applicable inferences.
11 . The method of claim 8 , wherein the modification strategy includes training the given trainable software model including varying weighting of inputs to the given trainable software model based on interpretability data derived from the inferences associated with the given trainable software model, leading to the unexpected output and identifying inputs to the given trainable software model that skew the inference to produce the unexpected results.
12 . The method of claim 8 , wherein the modification, strategy includes reversion to a prior version of the given trainable software model, stored in vehicle memory, providing output closer to expectations determined based on the unexpected output compared to prior output of the prior version of the given trainable software model, observed and stored in vehicle memory.
13 . The method of claim 8 , further comprising providing interpretability reporting, based on identification of unexpectedly impactful input values of inputs identified by the configuration file for the given trainable software model as having above a defined threshold impact on the inferences resulting from the given trainable software model.
14 . The method of claim 8 , wherein the expected inference values used in the identification of unexpected output are further based on vehicle-stored historical output values previously generated by the given trainable software model or a prior version of the given trainable software model.
15 . A non-transitory computer-readable storage medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform a method comprising:
subscribing to one or more inference topics to which inferences are published on behalf of trainable software models executing in a vehicle computing environment;
receiving inferences from the topics as the inferences are published to the topics;
associating the inferences with one or more trainable software models to be monitored;
identifying instances of unexpected output based on comparison of received inferences, associated with a given trainable software model to be monitored, to expected inference values identified in a configuration file, the expected inference values based at least in part on values provided from a remote server to the vehicle indicating historical output values previously generated by variants of the given trainable software model executing in other vehicles, stored in a vehicle memory, and associated with the given trainable software model; and
responsive to identifying the unexpected output, devising a modification strategy for the given trainable software model based on characteristics of the unexpected output.
16 . The storage medium of claim 15 , wherein the inference topics are identified based on information received responsive to a configuration file, associated with the given trainable software model to be monitored, requesting loading of the given trainable software model for execution and wherein the associating further comprises associating the inferences based on a configuration file associated with each given trainable software model to be monitored, indicating applicable inferences.
17 . The storage medium of claim 15 , wherein the modification strategy includes at least one of training the given trainable software model including varying weighting of inputs to the given trainable software model based on interpretability data derived from the inferences associated with the given trainable software model, leading to the unexpected output and identifying inputs to the given trainable software model that skew the inference to produce the unexpected results or reversion to a prior version of the given trainable software model, stored in vehicle memory, providing output closer to expectations determined based on the unexpected output compared to prior output of the prior version of the given trainable software model, observed and stored in vehicle memory.
18 . The storage medium of claim 15 , the method further comprising providing interpretability reporting, based on identification of unexpectedly impactful input values of inputs identified by the configuration file for the given trainable software model as having above a defined threshold impact on the inferences resulting from the given trainable software model.