https://www.st.com/content/st_com/en/campaigns/nanoedgeaistudio.html
Part2 - NanoEdge AI Studio:
Hi, can the model be saved locally to allow power failure on the sensing device and restart with known model?
> Dynamic model knowledge (Anomaly Detection) is allocated in RAM and it is possible to backup the model in the non-volatile memory as: EEPROM, Flash, etc. Please see: https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Library_for_anomaly_detection_(AD)#Backing_up_and_restoring_the_library_knowledge
Does the library support fixed point arithmetic for speeding up processing?
> We don't support it yet, but R&D is working on it.
Could NanoEdge AI studio be used if we are not able to generate "abnormal" signals?
> Yes, it is possible, please see https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Library_for_1-class_classification_(1CC)
How long will the selection take in a real application, for example a pump?
> In practice it varies between several minutes to few hours, please see: https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_examples
With the NanoEdge AI is it possible to have several types of fault pattern detection? not just one but a list of them we would need to recognize?
> Yes, it is possible using n-Class Classification or Multi-library use case, please see: https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Library_for_n-class_classification_(nCC)
https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Studio#Multi-library
What happens if we make learn at start during an abnormal situation?
> In this case, dynamic model learns abnormal condition as regular one.
Can we use this library with other sensors like air quality?
> NEAI Studio is sensor agnostic then any sensor stream could be considered (considering output data rate as limitation)
Could you please summaries how the N-Class works - classifying multiple anomaly types?
> Please see: https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Library_for_n-class_classification_(nCC)
Can we find any finished code examples for STWINKIT1B to do exactly what was demonstrated?
> Please see: https://www.st.com/en/embedded-software/fp-ai-monitor1.html
Is neural network learning possible on edge?
> Yes, it is possible.
Can you tell us anything more about the algorithms used for anomaly detection and classification?
> Please see: https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Library_for_anomaly_detection_(AD)
https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Library_for_n-class_classification_(nCC)
For training, is it necessary to use STM32 development boards or can the data acquisition be done on boards ready for production?
> Data collection process is hardware agnostic, the only point to consider is that STM32 MCU is a target hardware for NEAI library deployment.
Is Nano Edge AI available for BlueNRG-M2?
> NEAI library can be deployed on the top of STM32 MCUs.
Are your models proprietary?
> Yes, a part of the models is proprietary and patented by ST R&D.
After a Reset, the learning is lost or stored?
> NEAI library knowledge as dynamic one (Anomaly Detection) is stored in RAM then is lost after MCU power cycle however it is possible to backup the knowledge, please see: https://wiki.st.com/stm32mcu/wiki/AI:NanoEdge_AI_Library_for_anomaly_detection_(AD)#Backing_up_and_restoring_the_library_knowledge
How would you prevent overfitting when retraining on edge in a slightly different scenario than during model selection?
> It is important to emphasize that it is not retrained, but simply trained. benchmark = no training in anomaly detection, only lib selection. Overfitting during benchmark is prevented by the design of the process itself (multiple random subsets created, for training [lib selection], cross validation, testing). Overfitting also prevented through thorough emulator testing using new data. And in the target device, by learning the appropriate/recommended amounts of signals.
How would you just deploy a fully trained model, with no need for training every time? Can we have a fully compiled version of the firmware ready to be deployed?
> Yes, it is possible for Classification (n class and one class) and Extrapolation use cases as static models. Considering Anomaly Detection, it is possible to backup the knowledge and then clone to similar devices.
