This might be my shortest blog yet. Artificial intelligence comes by many names including machine learning. Systems that understand hand writing are not referred to as AI but rather optical character recognition systems. Deep learning on the other hand is an AI technique. AI can be part of many systems but is not an end in itself.
Anyway, here is the key guidance from the generic functional safety standard IEC 61508-3.
The use of AI is not recommended at any SIL level greater than SIL 1. At SIL 1 it is neither recommended or not recommended. For guidance the definition of not recommended is given below from Annex A of IEC 61508-3:2010.
One of the main objections to AI is that it is overly complex. Functional safety loves simplicity. To quote the book “Software for dependable systems”. Actually, I searched the book but couldn’t find the quote. I googled it and found it again in the book “Code Complete” and attributed to C. A. R. Hoare – “There are two ways of constructing a software design: one way it to make it so simple there are obviously no deficiencies, and the other is to make it so complicated there are no obvious deficiencies”. I see their point, when you consider that a deep learning algorithm might need to crash a car into a tree 50,000 times before it figures out it is a bad idea. A kid on a tricycle generally figures it about after the first or second crash. Non-determinism is hard to accept for safety. To give a second quote from the above book – “essential that developers are familiar with best practices and deviate from them only for good reasons”.
I can find no mention on AI in the automotive functional safety standard ISO 26262 and therefore in theory the guidance for automotive would fall back to IEC 61508. Yet there appears to be widespread use of AI within new automotive technology. I haven’t yet read all of ISO 26262 revision 2 (expected release 2018) but I must discuss this with my automotive functional safety colleagues within Analog Devices. Perhaps AI is only proposed for driver assist as opposed to safety applications. Perhaps it will be somehow covered by the new SOTIF standard (safety of intended functionality). I feel the benefits of AI may become so great that the above guidance may have to change and in fact IEC 61508-7:2010 clause C.3.9 offers such hope when it states, “supervisory actions may be supported by artificial intelligence (AI) based systems in a very efficient way in diverse channels of a system”.
Today’s video selection had a lot of possibilities. I went with the SpaceX heavy launch and side booster landing which took place the week I was writing this article. Elon Musk is one of the people who actively warn about the dangers of AI (Google it for a long list of references). I should probably have gone with something like HAL from 2001 a space odyssey but instead I selected the talkie toaster from Red Dwarf. Perhaps not what Elon was warning about but who knows perhaps he does watch "Red Dwarf" after all he obviously watched "The Hitchhikers guide to the Galaxy".
For next time, the discussion will be on the functional safety and security.