A little bit of a few things.

Wow, more than a month since my last post. I’ve been a bit braindead as I try to finish up my thesis. I’m going to be continuing in my current lab next year, only I’ll be doing my PhD. So that’s pretty exciting.

I’ve seen a couple interesting talks in the past while, and figured I’d post on a couple of them, briefly.

Imaging and intelligence: The presenter is a big g theorist, which given that this appears to be the current consensus isn’t such a big deal, but see a list of discussion links here and an excellent critique of the origins of g-factor theory. I need to look up my notes, but essentially the presenter’s group used various tasks and controls to find activation via fMRI for what they claim to be g in various brain regions, including that magical region, the IFG.

I went to a workshop/seminar on rhythm in music and speech. It was neat, and there was some discussion of comparing musical compositions to speech. The main problem I saw was that a lot of the studies seemed to compare compositions to spontaneous speech, or novel reading aloud. The two aren’t really comparable. A better comparison would be a composition with a composed and practised speech, or improvisational jazz with spontaneous speech. There was a talk after, where they discussed the use of tabla drumming as a speech-like code. The cooler part was a discussion of language-specific differences in supposedly universal aspects of acoustic perception. Apparently, when humans hear a continuous stream of tones which proceeds long-short-long-short-long-short-long-short… we order it short-long, rather than long-short. And when we hear a continuous high-low-high-low-high-low… we parse it as high-low rather than low-high. And this is thought to reflect inherent biases of auditory processing. But the speaker’s group found that this is true in very young Japanese and Canadian infants. But by the time phonemic pruning occurs, infants begin to acquire language-specific biases, and so Japanese infants will hear long-short and sometimes low-high.

Then there was a cool talk on emotion and speech. This one got me thinking, since it dealt with the minimum amount of an utterance you need to hear before you can accurately identify the emotion being expressed, assuming the words themselves are neutral, or if the sentence uses pseudowords. But I’ve been thinking that interjections like “Ugh!” or “Argh!” bypass this to some extent, and may also be culture- or language-specific. It’s part of a broader interest I’ve been developing in socially stereotyped behaviours, which posses a kind of social exemplar. We can all picture, and even imitate, a stereotypical sneeze, and I think this cultural idea of a ‘sneeze’ actually shapes how people sneeze. The same for laughter, stubbing our toes, wiping our eyes when crying… I think this idea of stereotyped behaviours is important, and could be studied in the context of verbal or gesture+speech communication as an efficient communication code or cipher. I need to bone up on my ethology.

I went to a comps presentation on universal grammar and connectionist accounts of language transfer. The speaker pointed out that neither camp makes sufficiently different predictions here for either to be falsified. I’m still sort of amazed that there are still UG people around, but I guess the theory has some explanatory power.

Monday there was a talk on detecting white matter activation in fMRI. I’ll explain sometime why this is generally treated as improbable, but essentially while there’s a good explanation for why we see changes in blood oxygenation levels coupled with grey matter activity, there’s no real explanation for what it would mean to see the same changes associated with white matter.

Tuesday I went to CRIUGM to see a talk on machine learning applications of multivariate pattern analysis in resting-state fMRI (where the participant does nothing except be scanned) and real-time fMRI. I need to go through my notes and do something more thorough, but it was pretty exciting, and showed some ways that we might eventually be able to combine fMRI with real-time conversations, and note relevant activations with specific parts of the discourse.

Alright. That’s enough for now. Things are coming together, so hopefully I’ll get back to posting more regularly.

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.

Direct realism and perception

So there’s this thing called ‘direct realism’, which is a kind of approach to the study of perception. It’s fairly involved, but essentially what it means is that the object of our perceptions, known as a ‘percept’, is the distal (causal) stimulus, not the proximal (sensory) stimulus.

Let me give an example. When you hold a cold can of pop, you don’t perceive the actual specific tactile input from your hand. Sure, you can sense that you are holding something that is cold, cylindrical, flexible, and with a bit of heft, but you can’t not perceive that you’re holding a can. Even just saying “I’m holding a metal object” means that rather than sorting out all the specific proximal signals that specify a metal object, you simply perceive the distal cause of your perceptions: that is, the bar. If you want to get really picky, even “cold” isn’t perceived as a transfer of kinetic energy from your body to the can, but as a thing in and of itself.

It’s important here to note that direct realism doesn’t claim that those proximal signals are not received by your body and brain, but that you, whatever that is, don’t perceive those signals, just their cause. If I ask a question, you don’t (necessarily) explicitly perceive the rise in the fundamental frequency of my auditory signal; instead, you perceive that I’ve asked a question.

Thus, direct realism claims that there is a strong, unavoidable top-down effect on perception. A bottom-up effect is one which proceeds from pure sensory input, such as sound waves or light beams or skin deformations; top-down effects rely on cognition and knowledge. By my understanding of direct realism, without any knowledge of a car door, for example, you might hear one slamming as a house door, or the loud thump of a wide object falling. But you still, despite not knowing what a car door is, would not perceive the structured compression of air waves. The sound would be ‘tagged’ to a hypothetical cause. Our experience thus structures out perception.

Direct realism is opposed to ‘indirect realism’ or ‘representationalism, which claims that we reconstruct an object as an internal mental representation, using information extracted from sensory reception. Indirect realism claims we have no direct knowledge of the external world, only internal representations of the world. Direct realism posits instead that while the signals from the external world do indeed travel through our sensory organs into our brain, going through neurophysical transformations, but that this is only a path for the signals to follow rather than the signal stopping and being constructed internally.

While this can end up being a pretty philosophical debate, it does have some real effects. If I’m studying speech perception (to pick an entirely non-random example), direct realism states that the percept is what’s being spoken; in one interpretation, the percept would be the articulatory gestures which structure the auditory signal. But from an indirect realist position, we might want to focus on the auditory signal itself, and how it is deconstructed and rebuilt into an internal representation of the words being spoken.

I’m pretty sympathetic to the direct realist view, myself. For one thing it posits a direct relationship to the external world, which suits my atheist/materialist/embodied-consciousness-ist views. It also supports a connectionist view of cognition. If we do not construct mental representations but simply perceive the distal object through a complex pathway, a connectionist network is much better suited than the sort of organization posited by symbolist theories of cognition.

A view of the brain.

I’m going to give this short post thing a try, without just posting a video.

So there’s a lot of different ideas about the brain. One of the more popular could be called “localized modularity,” and it’s been around since at least the late 70s, and especially since Jerry Fodor published Modularity of Mind. In it he describes the requirements of a modular system, which is how he sees the brain.

1. Domain specificity, modules only operate on certain kinds of inputs—they are specialised
2. Informational encapsulation, modules need not refer to other psychological systems in order to operate
3. Obligatory firing, modules process in a mandatory manner
4. Fast speed, probably due to the fact that they are encapsulated (thereby needing only to consult a restricted database) and mandatory (time need not be wasted in determining whether or not to process incoming input)
5. Shallow outputs, the output of modules is very simple
6. Limited accessibility
7. Characteristic ontogeny, there is a regularity of development
8. Fixed neural architecture.

Now, I have varying views on these. Bear in mind that I’m at a stage where I’d have trouble seriously defending my positions. But I have them none the less, so here goes.

1. This idea seems suspect to me. Why would it make sense for brains to have specific regions devoted only to language, or to tasting only food, or to recognizing faces? The fusiform face area is supposedly specific to facial recognition in normal orientation, but there’s evidence suggesting it’s really involved in processing any categorical visual information on familiar objects.

2. This just sounds like nonsense. Language doesn’t need to be informed by motor systems for producing speech? We don’t integrate multi-sensory information when viewing a scene? We don’t need to reference our motor system when perceiving ballistics? There’s plenty of evidence that many brain functions draw on a variety of functional regions for processing, not just a single encapsulated module. These days we should be thinking about broad, parallel networks, not single processing units.

3. I’m a bit softer on point 3. I think there are a lot of mandatory processes, both low- and high-level. I think the direct realists make a good point when they suggest that while we can choose to attend to proximal stimuli (cool temperature, smoothness, rigidity) we are obliged to attend to distal stimuli (a can of pop).

On the other hand, we can exercise top-down control of perception as well. This means that we may be able to exert conscious control over certain mental processes – whether Fodor would claim that those portions simply aren’t modules is unknown to me.

4. Again, I feel that this is true for some things and not for others. Perceptual functions might be more obligatory than motor functions, but I’d definitely argue they aren’t strictly encapsulated either.

5. I’d actually agree that network outputs tend to be simple, and that more complex effects are due to multiple network outcomes overlapping in time.

6. I do think we have limited access to mental processes, so I’ll let this one stand.

7. Again, I don’t have a huge problem with this, though I’d caution against genetic interpretations of this for neural development. I think the common brain development patterns (in terms of broad functional localization) is due to a combination of genetics, epigenetics, and massed Hebbian connective wiring.

8. No, no, and no. I can’t imagine anyone defends this one these days. While I think it’s still a bit of a buzzword, our current understanding of neuroplasticity really doesn’t allow us to support any model that includes fixed neural architecture. Use it or lose it seems to be the rule of the game, at least within certain broad bounds.

I’ve given some criticisms that might seem to accept the modular model of the brain, so let me explain briefly how I see things.

I think brain functions are conducted from broadly localized functional regions, made up of networks of non-domain-specific computational units. I think Broca’s area is involved in general sequencing functions, not just linguistic syntax. I think the pre-SMA is involved in new action sequencing, and the SMA proper is involved in voluntary execution of learned action sequences. I think the hippocampus is involved in long-term episodic memory storage.

I do not think that there is any true double-dissociation for precise, domain-specific functions. I think anyone who expects to find anything other than broad functional associations, using our current level of technology, is in for disappointment. And I think if we are able to locate domain-specific functional networks, that the neurons involved will be mixed in among neurons involved in other related functions.

But I could be wrong.

Mirror Neurons

Wow, it’s been a while. I’ve spent the past month and a bit arranging to move to a new lab, in an entirely different field of neuroscience, so I’ve been somewhat preoccupied.

I’m now studying speech perception, part of the field of neurolinguistics, and it’s proving to be quite interesting.

Let’s talk about the brain. I know it’s a pretty broad topic, but I’ll narrow it down.

Sensory information, say auditory and visual information, enters our body via sense organs (the ears and eyes in this case). From there it travels to the primary sensory regions (the temporal and occipital lobes). Bear in mind that this is all pretty rough – there’s a lot of detail I’m leaving out.

So from these primary regions it travels to secondary or association regions. In particular, let’s focus on the primary somatosensory cortex, in the parietal lobe. This region was first investigated methodically by Dr. Wilder Penfield (and here’s a shout out to the MNI). This region is the location of the sensory homonculus, due to different parts of the body being ‘represented’ in specific regions of this area (if anyone knows the original source of this old bit of illustration please let me know).

Now, different parts of the somatosensory cortex are composed of different neurons, which respond to certain stimuli in certain ways.

1. Some respond purely to one form of sensory input – tactile, auditory, visual, etc.
2. Some respond to certain subtypes of input – expanding and contracting visual stimuli, or arm movements away from the body, or specific configurations of finger joints.
3. Some respond to multiple sensory inputs.
4. Some respond under specific combinations – when a tactile stimulus is visible, and these will respond less strongly the less visible is the stimulus. And
5. Some respond sympathetically to stimuli.

What does this mean? Well, some will fire both under a certain condition and when you see or hear others receive that input. If a monkey sees another monkey pick up a banana, neurons which typically fire when there is tactile feedback from the banana-gripping hand position will fire when seeing another monkey gripping the banana. Or neurons in the lip region will fire when they see another monkey eating a banana even though they themselves aren’t eating one.

These are mirror neurons, and they’re pretty important.

The somatosensory neurons, you see, don’t just sit there. They project to motor neurons in the motor cortex, usually to ones involved in carrying out the actions that the somatosensory neurons respond to. So when you see someone reeling in a fish, not only do the neurons which respond to that sensory input fire, so do the ones involved in firing when /you’re/ reeling in a fish, and the motor neurons involved in reeling in fish are also activated.

This is a huge deal. The main theory of how learning occurs in the brain is known as Hebbian learning or Hebbian synaptic plasticity (after Donald Hebb, another founder of modern neuroscience). It states, basically, that whenever a neuron is triggered by another neuron, the connection between the two gets stronger, and the triggered neuron is more likely to fire when it’s stimulated by the triggering neuron in the future.

If the neurons involved in certain actions fire when you just observe those actions, it means those neural pathways can be formed purely through observation.

Now just sit back a moment and think about this. What are the implications for this when it comes to learning complex but evolutionarily useful abilities? Such as speech, for example.

There’s a lot more to it than this, but I’ll leave off for now.