Strategy for ISO26262 Certification of a machine learning based Autonomous driving system.

Removing outliers in Machine Learning Inference is a lot harder than it looks, and this is going to be the key to certify.

To Simplify greatly, if you put 2 Neuronal networks next to each other, on a graph, you get something like this in Figure 1.

A quick gaussien distribution will allow you to find the outliers ... it doesn't mean you can remove that point, it means it is not mathematically in the usual space of answer for your ML network.

In figure 2, you get the obvious taking care of. With 3 of those Neuronal network, you will be able to start extracting "the Truth".

The key to certification of your AD system will be to take multiple neuronal networks, and demonstrate to the regulators that you are having exclusively diverse outliers across your networks... That is hard to do ...

Outliers in Figure 3 are very similar, this is unlikely to help you to get this solved. You want the outliers of your networks to be in a different dimension to be able to prove your case.

For example, Yolo 8 is doing detections that are boxes, and resolve the localisation of the object in 2D.

To oppose to the 2D nature of Yolo8, you can use a post fusion inference to change the number of dimensions, and when you do so, now, what you have left is to demonstrate that your outliers are from a different dimension space, they mathematically can not be the same, using the theory of core and graphs. ( Core (graph theory) - Wikipedia )

Et voila, you have a way to certify a machine learning system, mathematically robust.

Of course, you would be inspired to add a non-machine learning system on the top of this all thing to be able be certain of your decision, I call this a certainty system (As seen in Figure 4).

What you have learn by reading this is that you will not be able to certify only with one gigantic neuronal network, as your outliers will not have the right properties allowing them to be insulated better than with a Gaussien. You will need an hardware capable of doing those comparaisons, and extract the outliers from multiple pipelines while doing this efficiently to do not cost you too much power.

Stay tuned, more coming ...

Francois Piednoel.