From AI to IA: an anatomy of the digital brain

AI is generally acknowledged as a contributor to and a creator of added value. But some view AI as a possible liability to organizations, to ourselves and even to mankind in general. As an AI aficionado for many years, I was given the opportunity during and& Leuven to shed a light on AI: what is it (not), and how can it help us, now and in the future?

Let’s start with what AI is not: it is not an omniscient robot who can do everything better than humans and eventually will decide to get rid of us. Firstly, the artificial intelligence - even though it has made immense progress - is nowhere near our human intelligence. Even though machines can by now beat world champions at chess and (the Asian game) go, it is still struggling with basic tasks such as recognizing a cat when it is displayed in a less traditional pose.

But it is not surprising to see people react defensively when confronted with a new disruptive technology. The same happened about a century ago when electricity started to make its way into our houses with electric light. Electricity vendors had to assure customers repeatedly that electric light would not be harmful to their health or the soundness of their sleep.

We should not forget that AI has been around for many decades. I have devoted my university graduation paper to a neural network system predicting the future energy consumption several years ago, but I have then moved in other directions because the applicability in real life was rather limited. In the meantime, the required technology has improved immensely and the need for machine assistance has grown at about the same speed, due to the increasing amount and diversity of data created every second.

Machine learning explained

One of the most interesting aspects of AI is machine learning. In a nutshell, this means that you can teach machines what output to expect when given a specific input. In a first stage, we will provide the machine with lots of input and output data, allowing the machine to create an algorithm which will produce the right output based on the provided input. In a next phase, we let the machine predict the output based on new input. The generated output will be compared with the actual result, and will be either validated or refuted. Based on this learning phase, the accuracy of predictions will increase gradually. Eventually, the machine will be ready to predict the future (=output) based on current data (=input). In a next stage, these so-called ‘magical insights’ can be reproduced and turned into automated decisions based on these insights.

Unconscious bias

With the computing power available today, many forms of AI become increasingly reliable: image and voice recognition, facial recognition, natural language processing, speech recognition, etc. But this has not led us to the Brave New World we would like to reach. Many times, we still end up with an unwanted result. A child using the sleeping mom’s fingerprints to buy Pokemon gifts, Alexa starting up a porn video instead of the kids song your son has asked for, the Asian man whose picture for a photo ID was refused because “subject’s eyes are closed”: these are just a few examples of how limited AI can be at times.

One of the pitfalls that is illustrated here, is the unconscious bias leading to wrong conclusions: when an Asian face is refused because eyes are closed, this problem has been created by using Caucasian faces as input material. This will be one of the biggest challenges in the adoption of AI: making sure that the ‘intelligence’ we create is a reflection of the actual world. In order to do so, we need the teams in charge of creating AI applications to be as diverse as possible.

FAT algorithms

One way of ensuring our AI applications produce the best possible results, is by making sure our algorithms are FAT: Fair, Accountable and Transparent. Roughly put, this means that we not only strive to create algorithms that are a true and fair reflection of the actual world, but also that we provide insight in the reasoning behind algorithms, and in the input data and logic used to create the algorithms.

From AI to IA

We are still a long way from the independently acting robots, androids and other machines we know from science fiction movies and series. There are so many brain functions that machines cannot and will not match for many years. But we do see an evolution from the state of artificial intelligence today towards a situation where machines become a logical extension to our own intelligence. That is what we mean with IA: Intelligence Amplification. The important ‘detail’ to notice: IA starts from our own intelligence and uses computing power to amplify our capabilities.

A very important detail, I might add, because that is how we ensure the general acceptance and succes of digital intelligence: by incapsulating the required ethics in the logic we create, and by training our systems as we would train our children: not only showing what is possible, but also what is desired and unwanted behavior and thinking. If we manage to achieve all that, we can look forward to a world where machines and humans join forces in making our organizations - and lives - better.