How intelligent is it?
In all of the demos, the A.I. system:
figures out for itself what its inputs mean, and what its outputs do, and how these relate to the physics of the environment it is in.
figures out for itself how to exploit the environment dynamics to get to a desired objective.
is very general.
In addition to these, Demos 2 and 3 make some non-Markovian deductions about their environment:
The spacecraft devises a strategy on how to direct its sense (the scanner).
It remembers scan results and combines them to build some sort of internal representation of its environment (NB: with only one scanner, to obtain the gradient you need more than one scan result).
It then uses this internal representation of its environment, and combines this information with its learned ability of how to exploit the environment dynamics, to fly towards its goal.
Meta-learning has occurred (see below for explanation).
I believe Demos 2 and 3 have learnt to be functionally similar in intelligence to a very simple organism. Other more complex intelligences may have more exotic skills, such as advanced language or vision processing, but as we’re only comparing to a hypothetically simple organism, we can ignore these and just consider control systems. The above list of 7 items make some “check-box” criteria that I think any system should satisfy to be classified as intelligent as a very simple organism, and the above list (particularly the second group) separates this project from most other A.I. systems and definitions of intelligence at this simple level.
Note that these demos do include a vision system of kinds, i.e. the scanner, but it’s an extremely narrow field of view.
To be functionally similar in intelligence to a very simple organism, memory of past interactions with the environment seems certainly to be necessary. That would mean that for neural networks, a recurrent architecture would be necessary. This would discount Demo 1 from this definition of intelligence. Demo 1 just seems to have learnt a function that gives it instructions of what to do in each position, e.g. “Turn thrust on, if velocity is high and downwards, and space craft is pointing upwards”, etc. Demo 1 is entirely Markovian (Markovian means that everything the spacecraft needs to know at any instant is given entirely by the state variables: position, velocity, orientation and fuel), and similarly any Markovian system would be discounted from this definition of intelligence.
Also importantly, in the later demos, some degree of meta-learning has been achieved. The neural network learning algorithm has produced a neural network that itself learns about its environment as it flies the spacecraft, i.e. it also learns a bit during each flight. This is a powerful combination: something that has learnt how to learn. With meta-learning like this, the system could in theory adapt to, or learn from scratch, new scenarios and problems; or tackle any other problem of similar complexity.
Are there any other criteria for intelligence that should be met? I am only really interested in what is needed to be functionally equivalent to intelligent systems - I don’t want to get bogged down in considering what is needed for a system of this kind to be conscious or not. (In my opinion, any system simulated in computer software certainly will not be conscious. See the Chinese Room Thought Experiment for a great discussion on that.)
However having made these claims of intelligence, I think these demos only just check these boxes, and only once each, for these attributes of intelligence. I’m sure the simplest organisms are far more intelligent than these systems, because they can do multiple types of task easily. I’ve applied this machine learning process to the most complicated task I can get it to master - and here are the results. So there is no danger of it taking over the world, just yet.
Comparison to Hypothetically Simple Organisms
What is meant by a hypothetically simple organism?
I’m not a biologist but I’ve read even amoebas are surprisingly
sophisticated (e.g. see
here). Ants are clearly far more sophisticated than these
demos. So I guess a hypothetically simple organism means
somewhere between an amoeba and an ant.
This is what the
criteria for intelligence are meant to apply to. Are these
criteria sufficient for this level of intelligence or is there
something missing?
What about higher level
intelligences, such as that of a mouse? Is it just more of the
same? Do the intelligence criteria I’ve listed still
apply? The learning algorithms I’m using certainly don’t
scale up to the mouse-level, but maybe the intelligence criteria do?
The Optimised Function Criticism
Critics of the view that this system is intelligent could say that all these demos have produced is an optimised function. That is true, but there are several counters to this: Firstly the functions that have been optimised here were very tricky to optimise, so it’s still no mean feat. Secondly, the same criticism could be made of any hypothetically simple intelligent organism (their function is reproduction success); and that’s all I’m trying to compare it to. And finally: So what? If a function had been optimised to pass the Turing Test, then we wouldn't be complaining!
The Difficulty of Defining Intelligence
If you look on Wikipedia's page on intelligence, then you will see that there are multiple definitions for intelligence. Many of them are quite qualitative and unspecific. Out of the more specific ones, one that the Neuropilot project undeniably meets is that attributed to Sternberg and Salter as "goal oriented adaptive behaviour". However even this definition is not specific enough to avoid some unsatisfactory conclusions. For example if you interpret this definition very liberally, then a thermostat meets it. For a thermostat the "goal" is to maintain the temperature at the set level. The "adaptive behaviour" is switching the heating on or off.
This is why I've added the check-box criteria above that hopefully distinguishes the Neuropilot from lesser intelligences, including that of a thermostat.
The moving target difficulty for A.I.
It seems that before A.I. programmers have achieved a particular task, they (and everyone) may think that task requires "intelligence", but after they implement the A.I. system, and after people have studied the A.I. system properly, they say the A.I system isn't intelligent. In other words, the target to produce an “intelligent system” is a moving one.
For example, before chess computers were invented, people thought chess was a task that required "intelligence" to play well. After chess computers were successfully built, and people had looked at the algorithm used (Minimax), they said "well that's not intelligent. All it is doing is a brute force search through thousands of possible future move positions and combining it with some mathematical operations".
The same is true for tasks simpler than chess. For example take a simple case of linear regression, which finds a straight line of best fit on a scatter-graph of points (X,Y). Once you've found the straight line, you can use it to "predict" new values of Y for any new given X. If you didn't know how this was working, you might be tempted to say the system is intelligent. "Wow, this system has just observed some examples of a company's previous end-of-year profits, and then successfully predicted its next year's profit." But once you do know the method, of course, you laugh at the idea of it being intelligent. If we want to make the system even more intelligent, we could use "non-linear regression" (e.g. a "neural network"). The laughter should continue because it's just the same idea.
I thought the same for the Neuropilot demos. Before I created Demo 1, I thought this would surely be classed as an intelligent system. However soon after completing it, I wasn't so sure.
Embracing this moving target problem, I have identified a criterion for intelligence most people actually use. I'm calling it the "implosion criterion" for intelligence:
For a system to be intelligent, it is necessary that we do not understand how it works. If we ever figure out how it works, then it is no longer classed as intelligent.
I call it the “implosion criterion” because the status of any entity as intelligent implodes as soon as we understand it! This has unfortunate consequences for intelligence researchers and A.I. programmers: according to proponents of this view, an A.I. programmer will never achieve success. They may even tell you success is possible if problem “X” is solved, but they will change their mind when they see the solution.
What would happen if we finally understand human intelligence and then apply this criterion? We would cease to be intelligent at that moment!
©
Michael Fairbank. Last updated: