What does product health mean?

Hi Jordan,

Considering the domain is ML, I think they might refer to predictive maintenance for their devices, using machine learning. This is a fast growing area of application for ML.

Fabio

You bring up a good point about Machine Learning.

With regular electronics you always have the virus issue causing the device to malfunction. You also have physical component degradation with time.

With machine learning, it is possible to provide the device with data such that you can create a tremendous bias in its algorithm based upon its processing of "bad" data.

So the idea of product health now enters a domain where there can be an outside influence causing the device to "fail" to operate properly based upon it having absorbed false information.

As I have said before, I do not like this approach from a safety standpoint. In risk analysis you look for definitive cause and effect. With Machine Learning, you can never have confidence in the end result of any specific stimuli. The device will provide the output it deems best fits the stimuli based upon its "learning" experience.

I would think that these issues will create opportunities for advance diagnostics for devices that depend upon machine learning to assess the health of the accumulated inferences. Eventually someone will have to define the characteristics as to when a device reaches a condition in which it become so unreliable that it must be "wiped" and retaught before it can be trusted again.

So much for machines being better than people.

DAB

You bring up a good point about Machine Learning.

With regular electronics you always have the virus issue causing the device to malfunction. You also have physical component degradation with time.

With machine learning, it is possible to provide the device with data such that you can create a tremendous bias in its algorithm based upon its processing of "bad" data.

So the idea of product health now enters a domain where there can be an outside influence causing the device to "fail" to operate properly based upon it having absorbed false information.

As I have said before, I do not like this approach from a safety standpoint. In risk analysis you look for definitive cause and effect. With Machine Learning, you can never have confidence in the end result of any specific stimuli. The device will provide the output it deems best fits the stimuli based upon its "learning" experience.

I would think that these issues will create opportunities for advance diagnostics for devices that depend upon machine learning to assess the health of the accumulated inferences. Eventually someone will have to define the characteristics as to when a device reaches a condition in which it become so unreliable that it must be "wiped" and retaught before it can be trusted again.

So much for machines being better than people.

DAB

I think the intelligent autonomous machines still belong only to the Sci-Fi realm.

Having said that, I believe there is a big misconception about AI and ML. Like for any technology, it works great to solve a specific class of problems, but in general performs poorly, or fails miserably, when applied in the wrong context.

ML has shown great potential in solving all those problems that are too complex to be expressed in the classic algorithmic style, either because of the vast number of variables or the lack of "determinism" in the way a possible solution can be determined.

One prime example is object detection and recognition: this is a problem a human can easily solve, but to try and write an algorithm to detect and recognise object it is extremely complex. This is one area where the ML approach shows its value: using neural networks (CNN in particular) we are now able to successfully process images with many mixed objects.

As I mentioned in my previous comment, another field where interest is growing is predictive maintenance for devices.

There is no magic in this technology: it is simply the combination of big amount of data, which are used to train the neural network, and the morphology of the neural network used.

The training data are "post mortem" kind of data, where you can establish relationships between the device health parameters and the failure occurred. The training "morphs" the neural network to make it more receptive to the parameters monitored, and able to predict (in terms of probability) the event of failure before it occurs, just analysing how those parameters change in time (those are special neural network, that can trace sequences, called Recurrent Neural Networks - RNN).

The results are promising. But again, it is still just machines we are talking about...

Fabio