Local and easy Machine Learning Operations (MLOPs)
To ensure a continuous increase of the performance, new models are generated with the help of the operator. This is enabled by DALE the blazing fast deep learning algorithm powering iDetect-4.0.
Even, if at its fastest pace the cycle can be as short as 10 minutes, it can follow existing quality processes. Indeed, thanks to its user management mechanism, iDetect-4.0 can separate the role between the operator that can correct or enrich the dataset by relabelling erroneous predictions and the person in charge of quality who can validate the corrections, generate a new model and quality it. The qualification of the model is done by computing an accuracy score resulting from successive inferences on a test dataset. How the score is computed and how the test dataset is constituted are defined by the quality policy. We go deeper in details in further chapters. The qualification and the deployment of the new models can be done locally, without stopping the production.
About local history, models, and data of interest
As far as storage allows, iDetect-4.0 keeps the history of past records and predictions, allowing for digging up root causes of crisis.
Thanks to the metadata captured from the automaton stored along each record as well as quality score, it becomes easy to find clues of potential root causes of problems, being an overheat of the machine or a sudden succession of warnings in inference results.
Because iDetect-4.0 runs on the edge, it must keep the storage size under control. That’s why an automatic cleaning mechanism has been implemented. Its policy is set by users with administrator role (see corresponding chapter for further details).
One of the outcomes of this local MLOPs is that it allows to separate signals of interest from signals that do not contain relevant information.
Indeed, signals that are correctly predicted by the algorithm do not contain more information than the ones constituting the training dataset. Those which are not are certainly new defects or new nominal functioning mode of the machine. Those must be kept and could be sent to a cloud for long term archiving.
The operator feedback allows iDetect-4.0 to know when such an event occurs. It can then label the corresponding records. Up to the quality team to use those records in a new model after thoroughly validating them. At cleaning time, it becomes easy to separate between those records of interest, corrected and validated, from those that can be safely cleaned out.
To implement this mechanism, we rely on the deterministic nature of DALE the deep learning engine that powers iDetect-4.0. Most of time, with DALE, from a given dataset you will always get the same model. So, storing the dataset is equivalent to store a model. The advantage of storing a dataset instead of a model, is that instead of an opaque collection of weights you get records that you can listen back and reuse in backward analysis.