MaCro Philosophy
5-October-2017

Human vs Machine Intelligence

Be it wise or foolish, we are currently walking the path towards human-like AI

AI improves along the dimension that we use to measure it. If we use a human-inspired definition of intelligence to determine our measures of success, we should expect more human-like AI. If we use a machine-oriented definition of intelligence, we should expect less human-like AI. We are currently following the more human-like definition, and therefore heading down the path to human-like AI.

If, like me, you believe that it's important to ensure our intellectual successors can enjoy the human-like qualities of conscious self-reflection, then we should follow the human-oriented definition. If, on the other hand, you believe it would be an unforgivable mind crime to create conscious entities with the capacity for suffering (see my previous post for further discussion on this topic), then we should be careful to develop only machine-oriented intelligence. Either way, it is crucial to understand which components of intelligence are uniquely human-inspired, and which are more general aspects.

Machine-Oriented Intelligence

A common way of defining machine intelligence is to appeal to the number of problems that an agent or algorithm can solve. For example, Legg and Hutter give the following definition:

Intelligence measures an agent’s ability to [robustly] achieve goals in a wide range of environments.
This definition is further expanded in the article by defining goals in terms of reward signals given by the environment, and ordering the environments such that the simplest to specify (i.e. those that follow simple update rules and are the least ad hoc) are the most important. This is a great general starting point for a definition of machine intelligence, deliberately omitting more human-specific aspects. I will explore the human-like components of our understanding of intelligence with a beautiful example of intelligent problem solving.

The Domino Problem

Imagine you have a pool of dominoes and a 6x6 grid. Each domino covers two adjacent tiles and they can only be placed horizontally or vertically. You are tasked with placing the dominoes so that they completely cover the grid. Easy right? You can just place them all horizontally or all vertically. But now imagine the grid is shaped as below. Is it still possible to cover the grid?

One attempt at a solution could be as follows:

This doesn't work. There are two spaces left over that no domino can fit in. But maybe there is some way to rearrange the dominoes to get those two left-over spaces next to each other so that the final one could be added. Feel free to spend as much or as little time as you like before moving on to the solution.

Two Possible Solutions

Intelligent Problem Solving

Both the above are valid solutions. I hope one seems more intelligent than the other.

Imagine you come across three powerful machines, Brutus, Paris, and Marco, that have just been input the dominoes problem. Brutus takes the brute force approach, checks 5 million possibilities a second, and outputs the solution in just over 15 minutes. It is not clear exactly how Paris works. Paris is clearly doing a lot of processing, comparing this problem to other similar problems, solving simpler versions, and trying to work out generalisable properties of the problem space based on her prior knowledge. Just after 15 minutes Paris figures it out and outputs the parity solution. Meanwhile, Marco searches the internet for similar problems and fortunately finds an exact match. At exactly the same time as the others Marco also outputs the parity solution (the reason it took Marco so long was that he kept missing an annoying show/hide button).

How would you rank all three in terms of intelligence?

Brutus and Paris

There are many interesting ways to take this thought experiment. Perhaps the Parity solution seems more intelligent because it can now be applied to solve grids of all sizes. The Parity solution can almost instantly solve the problem on the left. However, it doesn't help at all on the problem on the right. On the other hand, Brutus can (eventually) solve both, and indeed any problem with a well-defined search tree, using exactly the same technique. It is not the generalisability of the approach that is the important difference.

Efficiency and Reuse

The real difference is not that the approach is generalisable, but that the solution is generalisable in such a way as to efficiently solve similar problems in the future. Paris can apply her solution again in the same, and similar, cases to save a lot of processing steps while Brutus has to check every possibility every time. Note that Paris needs to able to explain the solution (but not how it was arrived at) to be considered intelligent. A simple "no" answer would not be enough.

This efficiency criterion makes sense from an evolutionary perspective. It is a great survival skill to be able to efficiently transfer solutions from one problem to other problems of a similar nature. It seems to be a good candidate for an addition to the human-like definition that is not necessary in the pure machine version.

Paris and Marco

Things become even more interesting when exploring the difference between Paris and Marco. At first, I was tempted to say that Paris is more intelligent, and that Marco has somehow cheated. But the more I think about it, the more I value Marco's solution. Marco had to know what to search for, to filter out irrelevant information, and to endorse the correct solution. This is no easy task and it's hard to see how it could be achieved. However, considering Marco's approach under different parameters, such as his inability to solve a uniquely presented (unsearchable) version of a similar problem, it starts to feel less intelligent again.

Explainability

Interestingly, it seems to be that focussing on the inscrutability of the method that adds to the feeling of intelligence. The more I focus on the unexplainable components of Marco's algorithm, the more intelligent it seems. The more I feel like I understand them, the less intelligent. Similarly, part of the beauty of Paris's solution seems to be the way it comes out of nowhere. I first came across the domino problem over ten years ago, and something about the solution has remained memorable ever since. I think it's that leap you get when reading it. That "Aha!" moment when a solution pops out of your subconscious into conscious thought.

There is an obvious analogy here between unexplainable subcomponents producing an explainable solution, and theories under which consciousness arises from combinations of subconscious processing. We do not know exactly what makes "Aha!" moments appear as they do, but EEG data suggests:

"insight is influenced by multiple processes operating at varying time-scales". (Kounious and Beeman)

To make explicit the analogy, it seems important that the solution is fully understandable (conscious), but the process for arriving at it is not (subconscious processing). If I could consciously explain roughly how the solution was achieved, I would turn the subconscious processing into a potentially conscious process, and any seeming intelligence would disappear. Brutus requires a lot of processing work to get its solution, but I can explain easily how it does it. If Marco is using some unexplainable techniques to verify and select its search results, he seems intelligent. Conversely, if Marco is just returning and automatically endorsing the first result found, then ascribing intelligence feels like a stretch.

The explainability criterion is also evident in our ascription of intelligence to AI. When I first heard of a neural network classifying images based on their contents, it sounded intelligent. The more I learned about the algorithms involved and how they function, the less intelligent it seems. Now, such methods seem about as far from actual intelligence as you can get. This is the case even though I am none the wiser about the specifics of the behaviour of a single run of any of the related algorithms, or about what information any individual connection encodes about the inputs. Human-oriented intelligence seems to require that there are at least some components of the algorithm for which it is not possible to come up with even a rough description of how they work.

In future posts I will explore the idea that this unexplainability component could even be a key component of consciousness

Human-Oriented Intelligence

We have arrived at the following human version of the definition:

Intelligence measures an agent’s ability to robustly, via a method that's not fully explainable, achieve a wide range of goals via reaching an explainable solution strategy that can be applied to (re)solve similar problems more efficiently.
The two additional criteria seem to come directly from our evolutionary history (reuse) as conscious beings (unexplainable components).

Conclusions for AI Research

In terms of efficiency and reuse, practical AI research has obvious pushes in that direction. It may be important in the future to make a clear distinction between efficiency and reuse for direct cost effectiveness and more general reasons that could be linked to survival.

AI research is in an interesting state with respect to explainability. Deep neural networks notoriously act as black boxes with their exact behaviour at the level of individual connections hard to predict or follow. However, their general performance is well researched and explainability of solutions is a growing research area. As AI takes a more prominent role in medical, policing and autonomous vehicle applications, it becomes increasingly important for it to provide explainable solutions.

It looks like the general direction the AI is heading is towards the more human-like definition. Research on explainability will lead to solutions that are explainable, but advances in the technology and the incorporation of more and more subtle optimisations to algorithms will also leave fewer experts capable of understanding the details of the complete solution method. Those opposed to heading down the human-like route will have to choose whether it is better to attempt to reach full explainability, or avoid explainability at all.

Blog Homepage