Machine Learning

Release 10.11.0

Improvements in Machine Learning Engine

• Role Based Access Control:
  With this release, to use Cumulocity IoT Machine Learning it is mandatory that users are assigned to Machine Learning global roles or user groups. This was done to tighten the security of the product and avoid issues related to privilege escalation. Existing Machine Learning application users will see an Access is denied error if they are not yet part of Machine Learning global roles or user groups. Likewise, the Zementis microservice APIs will respond with 403 - Forbidden if the necessary permissions are not associated with the requests made by the clients. Refer to the Machine Learning guide for more details.

• Enhancements:
  • Extended the retention period for events created during training of clustering models enabling users to fetch information about clusters for a longer period of time.
  • Improvements were made in overall log patterns and log statements of the Machine Learning Engine microservices.
  • The Machine Learning application now shows all the error messages in the form of alerts to deliver a consistent user experience.
  • Text fields in the Machine Learning application now limits the input size to 200 characters. Checks enforced to ensure only safe characters are allowed.

• Upgrades:
  • Machine Learning web application upgraded from Angular version 8 to version 11.
  • Nyoka and Onnx microservices upgraded from Python version 3.6 to version 3.7.11.
  • Onnx microservice upgraded from onnx-runtime version 1.3.0 to version 1.5.2.
  • Zementis microservice upgraded to use Red Hat UBI8 as its base image.

• Bug fixes:

  • The HTTP response code for {{url}}/service/zementis/train/timeseries and {{url}}/service/zementis/train/clustering APIs updated from 200 - OK to 202 - Accepted.

  • JSON parser now throws a consistent 400 - Bad Request when an incorrect JSON is passed by the user. Earlier it was throwing 500 - Internal Server Error in some cases.

  • Fixed the issue where the scoring of an invalid file was throwing a plain text error causing the Machine Learning application to break.

  • Fixed the issue where appropriate validations were missing while scheduling a job as part of the scheduled processing feature.

  • Fixed the issue where in certain scenarios Zementis microservice would fail to start if it encountered errors while restoring existing artifacts.

Improvements in Machine Learning Workbench

• Enhancements:
  • AutoML improvements:
    • Unsupervised learning provides all the hyper-parameter options for anomaly detection while training isolation forests and one-class SVM models.
    • Supervised learning provides cross-validation accuracy in the training tasks.
  • Workflow improvements:
    • Workflow API accepts Test Size parameter.
    • {{url}}/service/mlw/tasks API computes training and testing accuracy.
  • Neural networks improvements:
    • Neural Network Designer now supports RNN layers (LSTM, GRU).
    • Introduced early stopping for model training.
  • Improve project commit/pull experience:
    • Project commits and pulls to/from the C8Y inventory runs as background tasks.
    • Control switches to the Projects page when commit is performed. Control remains on the Projects page when version pull is initiated.
    • In both cases, the UI will show a small loader to indicate that the task is in progress.
    • Project delete runs as a background task.
    • Users will have an option to navigate to the Tasks section and check the progress status.
  • Improved performance when working with large I/O bound tasks.

• Upgrades:
  • Machine Learning Workbench web application upgraded from Angular version 8 to version 11.
  • MLW microservice upgraded from Python version 3.6 to version 3.7.11.
  • MLW microservice upgraded from tpot version 0.9.6 to version 0.10.0.

• Bug fixes:
  • Fixed the issue related to AutoML getting stuck forever in ‘RUNNING’ status.
  • Fixed the issue with AutoML generation results within the tasks API.
  • Fixed the issue related to Python export from Jupyter Notebook.