https://www.st.com/content/st_com/en/campaigns/nanoedgeaistudio.html
Part3 - NanoEdge AI Studio:
Can NanoEdge AI undertake dynamic learning? for example swerve detection in a vehicle travelling over a random selection of road surfaces and weather?
> Yes, it is possible for Anomaly Detection use case, please see: https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Emulator_for_anomaly_detection_(AD)
Are your models all proprietary?
> Part of the models are proprietary and developed by ST R&D.
Is there any practical difference between an Anomaly detector and a binary classifier? (1 input with GOOD and BAD values)
> Anomaly Detection model output it is the similarity [%] and adding threshold (boundary) e.g., 90%, it is possible to implement binary classifier.
How can you reset the model once it has accumulated knowledge?
> It is possible to call relevant “neai_anomalydetection_init�? function, please see: https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Library_for_anomaly_detection_(AD)#Initialization
Is it possible to analyze and recognize patterns in audio signal with NanoEdge AI Studio?
>Yes, it is possible. I would suggest n-Class Classifier, please see: https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Library_for_n-class_classification_(nCC)
Is there any way to use collected data of a replaced pump after years of use - as train data for the new pump or change model? (Continuous learning) Or do we start from scratch after hardware replacement?
> Data collected can be used it is possible to label it correctly i.e., know which data is nominal (or within acceptable range) which is abnormal, and possible, get data associated to different anomaly types. Then, a new library can be generated using this data. Or users could try transferring the model learned on the old pump (dump/load) to new devices.
If we make a learn in an abnormal situation, will it be impossible to detect this abnormal situation?
> No, if we learn an abnormal situation, then it will not be possible to detect is as abnormal (it will become part of the knowledge = the norm)
Do we need to run each time an initialization phase?
> It is needed every MCU reset, or (for anomaly detection) if user wants to wipe all knowledge.
Can the NanoEdge AI be used for a low power project like for IOT sensor?
> Yes, it is possible.
How do I make sure I do not overbias one sample input by having recorded too little
of another sample input?
> input classes should be as balanced as possible, but if they're not, it is not necessarily an issue because the main metric "balanced accuracy" takes class imbalance into account. Please see: https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Studio#Designing_a_relevant_sampling_methodology
Which solution would be best for recognizing 3 different sounds?
> n-Class Classification
Can we access the generated model as raw data? I mean without any library.
> No, it is impossible.
How do you take external sensor inputs into NanoEdge AI? For example, 3 pressure sensors.
> Considering e.g., ST MEMS pressure sensor, it is I2C serial bus.
