AI Systems’ Security and Safety … A No-Brainer?

The June 2017 issue of IEEE Spectrum is devoted to comparing AI (artificial intelligence) systems to the human brain. It is a special report containing eight articles under the umbrella of “Can We Copy the Brain?” We must ask whether this a reasonable endeavor? Will it give us a better understanding as to how to build AI systems? At some level, the answer is “yes” when it comes to sensory and analytical behavior. But we are nowhere near understanding the critical roles of an obscure part of the brain called the cerebellum, wherein the keys to safe and secure cyber-physical systems may reside.

First, a little history. Many moons ago, when I was in my final year of a bachelor’s program in electrical engineering at Glasgow University, we had to come up with a report on one of a list of topics and then present them under competitive circumstances. I chose as my topic “Electricity in the Brain,” and I described the functions of the various segments of the brain in terms of electronic-systems.

We then presented our papers and mine received first prize, much to my surprise. I think that this result may have been partly due to my slowly unwrapping a brown-paper package containing a pickled brain, which the medical school lent me for the occasion (with the caveat to not leave it on the bus!), in front of the audience and judges. This brain thing was my father’s suggestion. I found it interesting that more than 40 years later, in February 2008, Jill Bolte Taylor raised gasps and applause from her audience by displaying a real human brain in a TED talk about “My stroke of insight,” which now boasts over 20 million views … see

Be that as it may, my presentation described the workings of the various components of the human brain including the neocortex, which embodies sensory, motor and cognitive functions, and (importantly) the cerebellum, which was assigned to merely coordinating motor activities and balance, but now is thought to have other important functions. Most, if not all AI researchers have ignored the functions of the cerebellum, even though its functions might prove to be the most important components of intelligent cyber-physical systems. The cerebellum could well be the seat of security and safety for the entire brain, which should make cybersecurity and safety engineers pay much more attention to its functionality.

The issue, which is not usually addressed by AI researchers, is that, while the cortex absorbs information and sends out commands to our various parts, one of the cerebellum’s roles is to monitor those commands to ensure that they are viable and not contradictory. An example that I used forty-odd years ago is as follows: When you want to raise your arm, the upper muscles must contract while the lower muscles relax. If your brain were to order both sets of muscles to contract at the same time, you’d have a real problem. Thus, the cerebellum acts as a moderator to ensure that the commands, which your regular brain transmits, are rational. In systems terms, it can be thought of as a negative feedback loop that pushes the overall system towards equilibrium.

But where are the cerebella in AI systems? The cerebellum was excluded from consideration in the IEEE Spectrum report. In fact, in Eliza Strickland’s introductory piece on “An Engineer’s Guide to the Brain,” the cerebellum is not even labelled! Is this why we have so many serious issues with AI systems’ security and safety? What we need, besides AI systems that emulate the brain’s sensory capabilities and motor actions, are systems that check on the reasonableness, including the security and safety, of these activities. Security and safety are absolutely necessary for systems that operate self-driving road vehicles and similar devices, and they are really important for all other cyber-physical systems.

New research, such “Cerebellar granule cells encode the expectation of reward” by M.J. Wagner et al, suggests that the functions of the cerebellum are greater than previously understood … see

Could there be other critical cerebellar functions that have not yet been discovered? It is just a hunch, but I would venture that, when more of the functions of the cerebellum are understood, we will want to incorporate those functions into AI systems, making them safer and more secure and reducing illogical and contradictory behaviors that AI systems might display.

Post a Comment

Your email is never published nor shared. Required fields are marked *