Original article was published by Fabio Veronese on Artificial Intelligence on Medium
In the exploding era of computing (ubiquitous, mobile, quantum or whatever suits you better) there’s still a sacred Graal we struggle to reach without success, even if we look closer every Moore’s law step we advance: Artificial General Intelligence (AGI).
Back in 2010 or so, in my days as Bioengineering MSc at University, I had my 10 minutes epiphany. I suddenly pictured that, some day, a reinforcement learning implementation general enough on a hardware powerful and beautiful enough might have led to a so-called strong artificial intelligence or artificial general intelligence. Indeed for those who do not chew machine learning at breakfast, this may look something really cool, but moving to a more concrete reality my realization was much more pragmatic.
Narrow Artificial Intelligence
In “traditional” Artificial Intelligence approaches, you pick for a task (the one you think it is worthy enough to be tackled) and put in place a supervised learning technique. This means you take some pre-computed couples of problem data and related results, you put an AI to sit and learn them, and you expect it to generalize so to solve correctly a new situation whose data it has never experienced. Just as a very abstract micro-example, if you want an AI to turn lights on when someone’s at home, you provide a set of people-at-home input data with corresponding lights-on outputs, as well as some nobody-at-home and lights-off. Conversely, Reinforcement Learning mimics animal (also human) learning process (you don’t say?), so the AI is equipped with a reward function: it starts with no precise strategy, but every time it turns light on with someone at home it is given a cookie, the same with the opposite condition (no people, no light), while wrong behaviors are discouraged (do the laundry?). Out of the metaphor, it may now be clear how the supervised learning may not look the right way to reach AGI, but a possible way to generate only narrow artificial intelligence: you cannot expect it to be taught anything and to generalize anything.
Coming back to my epiphany I was (not) quite surprised to learn, some-ten-years-later, one of the past days, that one of the most promising approach towards AGI is currently Deep Reinforcement Learning (DRL). I could not help it, but I started to think and fantasize of how could be it designed to spawn an AGI. And I found myself stuck right away in a three-fold lock.
Humans have instinct
We (and most of the animal world) come with a pre-installed software we simply refer to as instinct. This helps us to build rewarding strategies and to take steps when all options seem equal. The baby looking for mom’s breast to feed, the need to cry when we feel discomfort, the meaning of staring in each-others eyes are powerful driver. They are also the basis for other reinforcement strategies.
What should an AGI instinct be like? Can it exist without it?
Humans have a body and a brain with a physical structure
The way our brain perceives the world with sight, hearing, our movements tuning and coordination, and countless other features of the human marvel come thanks to hardware configurations. Our body moreover has its needs and we share them, they drive our choices and condition our thoughs.
How should we design the AGI hardware? Should it be a human-like or designed from scratch? Should it quack? Should it need to sleep?
Humans LEARN how to human
In our first 20 years (even 30 sometimes) of existence we basically have no idea of what we are doing. We experiment, we try to achieve independence, to learn how to learn. What do we expect from Mowgli, raised by apes or wolves? He cannot even wolf nor ape properly, but damn, he is intelligent indeed. We got countless generations of learning and teaching and evolution made it suited for us as a species.
Is it suited to teach an AGI how to human? How can we teach an AGI to be itself? Will we be able to recognize its genuine intelligence?