The Game of Intelligent Design

If you were shown a picture of a carpenter wielding a hammer and asked to describe the divide between intelligence and tool, you’d maybe say it’s where the hammer meets the hand. Similarly, if shown a picture of a person doing an internet search and asked the same question, some people might describe two divisions: between the hands and the computer’s keyboard, and between the search server and the software that performs operations that could arguably be considered to be intelligent. (People have gone as far as to consider Google the real-life SkyNet.)

In both cases my fictional questionee considered the person’s body to be on the intelligent side, but without too much more consideration it’s clear we need to revise a little. Are skin and bones intelligent? Most would agree no. Muscles? Probably not. Nerves? That’s a more interesting question, but still i’d say we know enough about how they work to be able to describe them as machines that carry out a well-defined task, even if the way in which they do it is fascinatingly complex.

So, where exactly is the boundary of intelligence? At what place can we say, “here be the smarts”? Is it the brain? In the previous paragraph i just said that nerves themselves are just machines – like really complicated hammers, and isn’t the brain just a collection of nerves? (Yes, i know, there’s lots of other stuff in there, but the neurons are still considered to be the most important for intelligence. Also, “nerve” is often just a synonym for a neuron’s axon, and a neuron is much more than that, but i think it’s close enough at the moment for getting the point across.)

But we know that intelligence is in there somewhere. And if the neurons themselves are not intelligent, it must be interaction between the neurons that does the trick. This is why IBM’s cat brain simulation is not so very interesting, but the Blue Brain Project‘s relatively modest effort is. The former is essentially an attempt at scale without realistic interaction, like 10 billion chattering teeth all wound up and set loose at the same time. It makes an impressive spectacle, but really only shows off IBM’s hardware. The latter is a realistic and meticulously detailed simulation of a single cortical column consisting of roughly 20 thousand neurons, from which the scientific findings will likely be enormously important. But, even in all its realism is anything about it intelligent (beyond the brains that it took to create it)? We won’t be able to say for sure until we understand what a cortical column does, but for the reason i explain below i’d suggest the answer is a qualified no.

Going higher up, can we say that a massively interactive system such as the occipital lobe—which takes up roughly a third of the cerebral cortex and is responsible for vision processing—is intelligent? Let me restate the question: would it be of much use on its own? For that matter, would any of the various brain areas be any good on their own? We know from clinical studies that the brain can compensate for the loss of multiple systems, sometimes seemly without negative effect, but that there is a limit. Usually sooner than later there is a noticeable loss of what we consider to be “with it”-ness. (Astute readers will note that an unfortunate person in a “conscious coma” with an intact brain will also fail this test, yet should still be considered intelligent. I’d agree, but suggest that this is failure of the machinery peripheral to the intelligence, not of the argument.)

Getting back to the original question of the boundary of intelligence, we can now see that such a boundary is an abstraction. It exists in function rather than form, and probably at a high level. The brain behaves intelligently through the interaction of its many high level systems. This result leads to many insights about how to recreate an intelligence within a computer – which i will leave to other posts to properly enumerate. For now we can note the following:

• If low-level systems do not manifest intelligence themselves, but rather just inform higher-level systems, it may be possible to abstract the low-level systems in order to simplify the problem. On the other hand, it’s possible—perhaps even likely due to evolutionary conservation—that the functioning of these low-level systems is very similar to how high-level systems work, and that for the most part the difference between the two is merely where they appear in a system hierarchy. My opinion is that there are both significant similarities between levels, and important differences.
• High-level system depend upon each other. Some development efforts currently in progress focus on the functioning of a specific area of the brain. For example, Numenta dedicates their effort entirely to discovering the “cortical algorithm”. (At least, that is the claim made in the book “On Intelligence” by founder Jeff Hawkins. In practice they abstract significantly even from that limited goal, as i determined from an extended discussion with Jeff and lead developer Dileep George.)
• No particular system is necessarily required for intelligence. This may at first appear to be a bold claim, but it’s not really. Skeptics might say that intelligence is impossible without perception, memory, association, pattern organization and recognition, etc. But in fact these attributes are ubiquitous at all levels, right down to the neurons themselves. Some mammalian brain areas such as the hippocampus appear to perform specific memory functions, but referring to it as a memory system is unhelpful; any neuron with plastic synaptic strengths, the ability to form new synaptic connections, and malleable second messenger systems—i.e. every inter-neuron in the brain—is a memory system. Certainly the hippocampus performs functions that aid in the detection, storing and recall of important patterns, but in doing so it is merely doing what every other brain system does: help the organism survive.

Google has once again stepped up the game with their announcement that users can now search by image. My first reaction when i read this was simply, “Wow”. Reading on, however, i came to the part about the “nascent nature of computer vision”. Anecdotes put the start of research into computer vision back to 1966 when an undergraduate student was directed to “solve the vision problem” over the course of the summer. Such is the degree to which we underestimate the complexity of the problem. About one third of our cortex is dedicated to processing information from our eyes in probably hundreds of distinct ways and then integrating all of the results together into what is consciously deemed to be an integrated whole. Nothing shows just how galactically complicated human vision is more than when over 40 years later one of the top technology companies on the planet calls our understanding of it “nascent”.

To be sure, some truly inspired work has been done in that time. Advances in some areas such as OCR and specialized machine vision have revolutionized productivity in certain industries. And depending on who you ask there are vision technologies peeking (no pun intended) over the horizon that will transform the world.

But did we start off on the wrong foot altogether? When someone decides to get into artificial intelligence, i’d say there’s probably more than a 50% chance that they’ll go into computer vision first. And almost without fail the first stop on the long, long vision path is image recognition. And why not? It seems simple enough to begin with a static image and process it until it concedes some kind of understanding of the scene. Certainly our computing technology most readily lends itself to this approach.

The problem is that this is not how human vision works, and naturally it is human-level vision that everyone is really after (just as it is a human-level intelligence that all AI researchers are after). Everyone has seen pictures of animals that are masters of camouflage, how they appear to be just another leaf or twig or rock or bump of sand until they suddenly move, and they are revealed. And those who read Jurassic Park remember how, if you just stayed still, the dinosaurs wouldn’t see you because they could only see motion; they couldn’t decipher a static scene. (Although the characters later – and unhappily – learned that the dinosaurs were in fact more cerebrally advanced.)

Could Michael Crichton have been right? Could it be that motion recognition is simpler than image recognition? It certainly seem plausible. Especially when an undeniable natural defense against predators is blending into your environment by being appropriately coloured and not moving. And if you consider how much simpler it would be to build a neuron circuit that detects visual change than one that detects arbitrary static forms, the argument becomes very convincing.

If we accept then that motion recognition came before image recognition then we can apply the old rule of evolutionary conservation and assume that the later is an advanced form of the former. Again, in practice we can see (again, no pun intended) how this may be the case. Detecting motion is better than not detecting anything, but only responding to relevant motion is better than responding to everything. And the better our assessments of relevancy, the better our survival.

But let us return now to how to build computer vision. Perhaps it is wrong to start off with image recognition. Perhaps the right approach is to build a computer vision that sees motion first, and through this can build up a repository of objects with a complete set of visual perspectives. Such a repository could then be used to decompose an image into the objects the scene contains. Given that the edges and gradients of an object move more or less together in a moving scene, it should be easier to associate features of a moving image than one that is static.

A significant problem with this approach is that our computer technology is not designed for it. I personally know of no computer languages in which time – even in an abstract sense – is a key feature. For example, if you wanted to model the potential of a neuron – how it changes based upon interactions with neurotransmitters and either tends back toward its resting state or fires an action potential – it would be entirely up to you to manage the changes that happen only due to time. Likewise, the release of adrenaline in an appropriate organism causes distinct changes, the effects of which diminish over time. Modeling such effects are completely up to the programmer. I believe this is why vision researchers tend to prefer static images. Our technology provides a simple way work with them. But very quickly the same simplicity ends up being a roadblock.

In GoiD i plan to introduce scripting features that will make time-based change simple to implement. Hopefully players will find this useful, and such concepts will spread beyond the game.

Anyone who has simulated the behaviour of insects in software has likely found that it’s not all that difficult to get some interesting things happening pretty quickly. A lot depends on how realistic the simulation is, of course – the more realistic, the more code that is needed to deal with real problems. For GoiD’s part the simulations are kept relatively simple so that the player can focus more on high level behaviours.

“Animals with bigger brains are not necessarily more intelligent” – Lars Chittka

As the article that this post takes its name from (http://www.sciencedaily.com/releases/2009/11/091117124009.htm) says, bigger animals need bigger brains because there is more to control, which is why animals of different sizes but similar intelligence have the same brain/body size ratio. Other things like highly-functional vision and fine motor control also take up huge amounts of brain space. (The occipital lobe – responsible for vision processing – is roughly a third of the cerebral cortex, and the cerebellum, which “merely” monitors voluntary movement, contains about half of all of the brain’s neurons.)

If what the article says is true, it may be possible to create GoiD scripts that behave something close to intelligently.

GoiD bring the concepts of artificial intelligence out of the university research departments and into the hands of anyone capable of writing a script. An important difference between typical AI development and GoiD, though, is GoiD’s tendency toward physical environments, as opposed to traditional AI’s focus on abstract environments (say, like playing chess). GoiD’s approach is reasonable because the only existence proof of intelligence is the human nervous system (i.e. your brain), which is in fact much like many of the other nervous systems found on the planet. And there’s one thing that all nervous systems on the planet can do, and it’s not playing Go or trading stocks: they can move.

“Deep Blue might be able to win at chess, but it wouldn’t know to come in from the rain.” – Marvin Minsky

This idea is not new, of course. Robotics labs all over the place create control systems in the spirit of “situational embodiment”, where the control system is “embodied” in a physical structure (the robot) and “situated” in an environment (e.g. a room), and this is a wonderful thing. The problem is that it is expensive and very time consuming. Computer simulations were the knee-jerk solution, but creating one that even remotely honours the laws of physics is remarkably tedious to write and slow to run. Also, no simulation is ever going to be exactly true to real life, so a robotic system will always eventually have to be tested in the real world before it can be said to actually work.

Still, there is a place for simulation in solving general problems and determining best practices. GoiD was created to allow people to explore the very large space of AGI problems in a way that is intended to be fun, and maybe a little competitive. With GoiD the effort of creating a simulation can be shared with everyone, so that as a group more actual AGI development can get done. Also, because players compete, GoiD creates a natural selection environment that – abstractly at least – is similar to how the human nervous system developed.

And if GoiD ends up also being fun to play, all the better.