Robots and machine intelligence.

So I went to a lecture on Friday by Dr. Roger Brockett, hosted by CIM and the Faculty of Science at McGill, entitled “What is an Intelligent Machine?” Dr. Brockett is the founder of the Harvard Robotics Lab, and an important figure in the field of robotics.

I really liked the lecture. Dr. Brockett talked a bit about the history of intelligence as a construct, both unitary and multiple, and I agree with his position that intelligence is more a collection of several (potentially over-lapping) abilities rather than a single underlying trait.

He talked about the Turing Test, and suggested that the test is not well-suited for non-linguistic intelligence. He pointed to wood-peckers, foxes, and other animals who display intelligence in specific domains at least equal to that of humans, and stated that we need to develop tests for robots that are more domain-general. In particular, he suggested tests which involve manipulating objects and solving problems in the environment.

I mostly agreed with him during his presentation, though I think neurons are less binary than he claims; despite the ‘all-or-nothing’ nature of the action potential, the summing of EPSPs and IPSPs seems analog to me. He talked quite a bit about how we have used analog and digital mechanics and math separately, and about how any modern approach to robot intelligence will have to utilize both modes.

But during the question period, we parted ways. Bearing in mind that Dr. Brockett is rather more educated in these matters than I, there were some points that I felt he is mistaken on.

1. Learning: Dr. Brockett was asked about the role learning would play in designing intelligent robots who would be able to problem-solve in their environment. I was quite surprised when he said there wasn’t one, and seemed to suggest that robots using his model would have all their required information to begin with.

This is a bit bizarre. He spent some time talking about how complicated even a small thing like shaking hands can be, but seems to think that this operation can be programmed into an intelligence. It would be much easier, to my mind, to create some simpler rules and then to allow the robot to explore its environment and to learn dynamic movement rules.

2. Neural networks. I think I’ve talked a bit about these before, and I think connectionist models are pretty awesome. But Dr. Brockett looks to be on the same side as Fodor, and argued that while neural networks are fine and good, especially for learning, what a complex robot intelligence need will require a more computationalist approach using formal symbol manipulation. I disagree entirely. I don’t think a computationalist, non-learning robot will be able to display the flexibility of behaviour that Dr. Brockett claims is the hallmark of intelligence.

3. Embodiment. Although Dr. Brockett didn’t use this term precisely, he spoke frequently about the need for robots to be able to move around and do things. Here I’m a bit torn. I don’t see why a robotic intelligence could not inhabit a virtual environment, or could not have sensory input/output without motor input/output.

That said, I’m a big fan of the idea that intelligence and particularly consciousness requires a certain degree of embodiement; but this is because I believe these things require learning in a changing environment, so I don’t quite understand Dr. Brockett’s focus on this if he feels there’s no need for learning in robot intelligence.

The last issue for me didn’t really come up explicitly, but I figured I’d talk about it anyways.

One of the reasons I focus on learning to much is it allows for more evolutionary strategies. We could design a quadrapedal robot, program all the complex business of walking into it, and then send it into an environment to move about.

But wouldn’t it be better if we created a virtual environment, allowed some random physical mutation, and made it so that success on these non-verbal intelligence tests Dr. Brockett was discussing caused the successful design to pass on its most recent configuration? Then we’d run the same tests again, and again, until we came out with a robot with physical and ‘mental’ adaptations well-suited to the virtual environment. We could then re-create these physically and test them again, using a learning neural network for the ‘brain’. We’d end up with something that worked, probably efficiently, and it would not necessarily be the shape we expected. It wouldn’t ‘think’ the way we expected.

This is a great way to approach robotics and intelligence, I think. Because of the random element involved in evolutionary processes, we do not have to pre-suppose a great deal beyond the starting conditions, and this allows for solutions that we would never have expected. Sure, you could spend ages designing a finger on a hand, but evolutionary robotics might come up with a suction appendage, or a flat sticky flexible ribbon, or who knows what else?

Admittedly, I’m overstating the case a bit. There are certainly strong limitations in my proposed methods, but I think they would prove out in the end to require less brute-force work and generate more interesting findings than the static, computationalist approach suggested in the lecture.

I should mention that Dr. Brockett took some shots at people who had concerns regarding quantum mechanics, free will, consciousness, etc. I generally appreciated this. While I obviously have a certain amount of interest in those topics, it was nice to see a pretty down-to-earth lecture on intelligence in machines.

The Measure of a Man

So, AI.

I’m going to be a sadly stereotypical geek and start off with a Star Trek reference. In my defence, I’ve only just started watching it, and will be moving onto the 3rd season as soon as I find a cheap / free way to get it that doesn’t involve hours upon hours of downloading.

In an episode in the second season, a Federation scientist requisitions Data (an android, essentially a unique being) with the plan of uploading Data’s data to another computer and then slowly deconstructing his positronic brain to fill in the gaps in his research.

There was a lot to like in this episode. Unlike a lot of the knee-jerk anti-technology/science attitudes in the show (which are always pretty jarring), Data doesn’t object to this in principal. Rather, he feels that the scientist’s research is not sufficiently advanced to ensure that Data’s consciousness will be preserved intact. A legitimate concern, I’m sure you’ll agree.

Eventually we get to the meat of the episode. Data is essentially requisitioned by the scientist, and a trial ensues to determine whether Data is Federation property, or if he is an independent being.

The good stuff continues. Riker is assigned to the prosecution, and executes a fairly emotional attack on Data by switching him off, playing up Data’s existence as a machine. The best part of the episode comes when Picard acts as the defense.

“Commander Riker has dramatically demonstrated to this court that Lieutenant Commander Data is a machine. Do we deny that? No, because it is not relevant – we too are machines, just machines of a different type.”

An excellent line. Picard further demands to know the scientist’s criteria for a sentient being. The scientist seems rather caught off-guard, a bit surprising given that anyone remotely involved in AI research gets caught up in questions of consciousness and sentience. He ends up defining sentience as self-awareness, intelligence, and consciousness.

Now, right there we have a problem. All three criteria over-lap quite a bit. Self-awareness in a basic sense is easy – as machines that have to move about and do things, both Data and humans are aware that they are beings distinct from their surroundings. In different sense, both types of machines are also capable of some degree of introspection or self-monitoring. This is a key evolutionary feature, as it allows one to judge internal changes due to external stimuli, and then to make compensatory changes.

Intelligence is even harder. The debate over just what it is, if there is one measure of several independent types of intelligence, and how to measure it has gone on pretty much since the word was first used in the modern sense. Intelligence tests at one point were almost trivia or general knowledge tests, and you lost IQ points for not knowing the name of US legislative bodies, or the years of the first world war. These days IQ tests rely much more on abstract reasoning, pattern recognition, and that sort of thing, but the debates continue (rightly, to my mind). I’d say intelligence as a criterion for sentience is pretty much useless, especially across species boundaries (in a setting where there are multiple ‘civilizations’ of different species).

Still, the episode once again uses a fairly trivial definition of ‘intelligence’, and Data rattles off some math and facts to mislead the viewing public as to just what makes people ‘smart’.

Now we come to the meat of the question. Consciousness. Now, anyone with even a passing familiarity with philosophy, AI, neuroscience, psychology, etc knows there is basically zero consensus about consciousness.

Can we measure it? Can we accurately report it? Which phenomena count as ‘conscious’, and which count as ‘unconscious’? Where is it located? Are animals conscious (or are the just intelligent)?

Picard goes back to one of my favourite arguments, functionalist that I am. He asks the scientist if he, Picard, is conscious. The scientist replies in the affirmative. He then asks the scientist to prove it. This puts the scientist in a bit of a bind, since Data is pretty clearly able to pass a Turing test. After all, he only lacks a bit of vocal intonation and knowledge of common social interaction, which are also found among humans.

This all leads to a huge facepalm on my part, as the adjudicator exclaims that the entire trial is to see if Data has a soul. Which makes me think that either the writer’s didn’t understand anything they’d written for the entire rest of the episode, or that they felt they had to dumb it down for the audience. Star Trek’s vitalistic tendencies are a whole other issue, though. You’d also think the Federation would have guidelines for dealing with any machine races they ran into, but no. It seems centuries of theorizing about full AI somehow never penetrated the Federation’s policy structure.

Well, this has ended up being more of an episode review than the essay on AI I’d intended. Clearly I’ll need to revisit the subject.

The take-home messages here: when defining your terms, don’t paint yourself into a corner. And don’t try to designate something as non-sentient that is able to carry on lengthy, unscripted conversations. We’ve already covered that. I wonder if we could call this the ‘Turing Trap’.

Functionalism fun.

So, functionalism. Us materialist-connectionist types tend to be big functionalists, even if connectionism includes a certain structuralist element. Since I think functionalism goes hand-in-hand with token physicalism (as opposed to type physicalism), this makes more sense than it may seem at first glance.

I probably owe some definitions here.

Functionalism is the philosophy that what things are or how they are structured matters less than what they do. Even for simple objects, functionalists believe function is more important than form. It doesn’t matter what a wrench is shaped like so long as you have a gripping lever of some kind. Birds and bats both fly. Brain and computers… well, that’s a big argument.

Structuralism is the philosophy that the structure and organization of things is what matters. This idea is mostly obsolete, though it does guide a lot of its successors. Humans, for example, rely heavily on the organization of our atoms. The same atoms in a pile are very different than when they’re human-shaped.

But that’s an interaction of functionalism and structuralism. We need to be in the right shape, but only insofar as our shape can preserve our function. If you replace my arm with a robotic arm, it doesn’t change my being a human or a person, though my structure has changed, because my function remains.

In fact, you could replace me piece by piece, even using foreign structures like treads instead of legs, and I would still be me and I would still be a person, because my function would not change. Even my brain, if it were transferred or translated to a sufficiently sophisticated computer, would be structurally different but not functionally different.

‘Substrate neutral’ is a phrase to remember here. Functionalism says it’s not what things are built of or built on, but what they do. If a computer chip does the same thing as your brainstem, it’s functionally a brainstem even if it’s silicon and gold and transistors, instead of carbon etc. and made of neurons.

I mentioned physicalism above. Scientists, being materialists and naturalists and monists, are physicalists (except for compartmentalizing religious scientists and physicists waxing philosophic and giving the rest of us no end of trouble explaining what they’re actually saying when they say ‘spiritual’).

Token physicalism states that specific mental effects, such as feeling a pinprick or imagining a face, are physical effects. It does not, however, specify the mechanism by which this happens, and is thus *ta-da* substrate neutral. Type physicalism, on the other hand, associates whole categories of mental effects with specific physical effects, such as pain reception via nociceptive nerves. This would suggests that animals with different physiologies, for example, could not experience the same mental effects.

I’ll devote a whole post one of these days to qualia – hopefully I’ll be able to keep a tight rein on my language.

Where am I going with this?

It seems like I’m drawing a contrast between structuralism and functionalism, with a weird diversion into species of physicalism. The fact is, functionalism and structuralism are related. Functionalists do not claim that types of function rely on specific structures (type physicalism). However, specific functions in specific agents, objects, or artifacts, do rely on their structure for their function (token physicalism).

I might be stretching the association a bit. I’m not a philosophy major, so I’ll try to sum up with how neuroscientists (or at least, the ones I agree with) see things.

1. Mental processes are physical. Anything that happens in our minds happens in our brains – mental events are physical events. This is a monist position, and it’s hard to defend anything else, unless you take Popper’s position (which is more of a social or informational monist position despite what he claims).

2. Mental processes are generally substrate neutral. That is, the specific materials and organization of the ‘brain’ matter less than the fact that it functions as a brain. Thus, there’s no reason why a computer could not have emotions, or be introspective about its motives.

3. Mental processes are specifically structure-dependent. That is, in a specific agent’s function would be disrupted by a disruption of its structure. This does not preclude the adaptation of new structures to existing or previous functions, but does mean that within an agent structure plays an important element. Computers and humans can both sleep, but humans would be definitely impaired if their thalamocortical relays were cut off.

To sum up: things are physical, even if they’re mental. Function matters most; it relies, however, on a general structure but not a specific structure. Different things can have the same function with different structures.