Neocorticalism and Its Weaknesses

[Blogger's note:  I went dark halfway through writing this.  Started it back in November, then got diverted into a self-imposed software project and lost interest for a while.  Not sure I'm back to blogging for real--we'll see.  No doubt you, my three readers, are incredibly excited!]

Just finished reading [well, last November] Ray Kurzweil's How to Create a Mind.  This has the usual Kurzweillian arguments about why all information technologies--including those associated with biology and neuroscience--grow at an exponential rate, irrespective of the actual technology base.  It extends his singularity books by going a bit deeper into how to build artificial cortical pattern recognizers, and makes the argument that a hierarchy of these modules pretty much constitutes the human mind and its consciousness.

Kurzweil references Jeff Hawkins's On Intelligence, one of my favorite books of all time.  Kurzweil takes issue with some of the details of Hawkins's model, but the two authors both agree on several central points:

• Both note the existence of a "cortical algorithm", where all areas of the neocortex work pretty much the same and project axons in pretty much the same way.  This means that patterns get recognized the same way everywhere, and project up to higher level pattern recognizers and down to excite, reinforce, or inhibit lower level patterns.
• Kurzweil thinks that consciousness is an emergent property of a big neocortex, while Hawkins thinks that consciousness is "what it feels like to have a cortex".
• Both of them spend a lot of time on input pattern recognition, but very little time on motor outputs.
• Both treat the thalamus and brain stem as I/O devices, with little to do with consciousness (although Kurzweil does note that the thalamus is essential to be conscious).
• Hawkins thinks that the secret sauce for complex cognition lies in the time-dependent behavior of cortical activation, where pattern #1 primes the downward links for pattern #2 to be activated if things go as the patterns expect them to in the immediate future.  Kurzweil is strangely silent on time-dependent behavior, which is weird, since speech recognition is so heavily time-biased.  I think he's glossing over some behavior of hidden Markov models in which I'm not expert.  He may also be slightly cagey about trade secrets.

I think that they're barking up the right tree with respect to the cortical algorithm, but they're both minimizing the importance of the older brain structures.  I have some trouble with this "neocorticalist" approach.  Here are my big objections:

• Neocorticalism can't really explain attention.  As I've said before, I think that attention is likely an incredibly ancient property and is intimately involved with consciousness.  I'm prepared to believe that human-style self-reflection and theory of mind might be new, emergent properties, but they can only emerge because of some repurposing of older mechanisms.
• Simple pattern matching doesn't get you very good motor performance.  I'm prepared to believe that the motor cortex is coordinating intentions and hifgh-level actions, but the whole business seems way too asynchronous to allow us to be catching balls or playing the piano.  The cerebellum is clearly involved, but I suspect that you need a way to make the neocortex semi-synchronous.  There's some evidence that the basal ganglia are involved at least in time perception; I'll bet we're going to discover that the same structures are "clocking" groups of neocortical patterns to coordinate activities in the motor cortex.
• One of the things that was intriguing about Hawkins's architecture is that it hinted that the learning process self-organized the cortical hierarchy so that novel or unlearned activities were first handled, somewhat hesitantly, at high levels, but the learning process pushed down the salient features into lower level learning, but no mechanism was described for how this happens.  Kurzweil pretty much ignored this, going so far as to posit some kind of central allocator of pattern recognizers.  There's something subtle going on here that pure neocorticalism can't capture.

Since this book was published, Kurweil has taken a high-level technologist position at Google, which has also bought, DNNresearch, Geoff Hinton's startup.  Deep learning, a term coined by Hinton, has been getting a lot of press recently.  It's pretty clear that Google has decided to pour money into this.  It's a problem that's amenable to implementation in cloud computing, so the fit is pretty good.

People implementing deep learning train each layer in a multi-layer network separately, using unsupervised feature detection.  ("Unsupervised" means that you don't tell the layer when it's done something right or wrong--you merely let it classify inputs as it sees fit.)  This still can't be quite what biological systems do, because they can figure out their own layering, which has to be imposed for deep learning.  My guess is that layering in humans is governed by chunks of the cortex that are genetically predisposed to accept axons from I/O-like areas of the thalamus and other parts of the brainstem.  They therefore learn to detect features associated with that kind of input and project out to a more amorphous set of regions, which can then combine multiple projections into novel features.

This still doesn't define a mechanism where layers compete with each other to identify features at the appropriate level of detail, but you can see how that might emerge with enough feedback between layers.

My guess is that we're going to discover that a huge amount of our cognition is dependent on a fairly arbitrary set of input regions mapping to another fairly arbitrary set of cross-regions, and so on.  But note that "arbitrary" doesn't mean that they're not genetically predetermined.  The stability of those mappings across most humans is what allows us to communicate with each other and what allows us to perceive most individuals as "sane".  One of the most interesting things about figuring out how all this hangs together is the possibility of producing entities that think radically differently from us but which still can extract insights about the universe that we're not optimized to perceive.

Things are going to move pretty quickly from here.  We aren't close to a strong AI using this kind of technology, but the Watson/Jeopardy thing shows that you can get an awful lot of interesting work out of something that you wouldn't necessarily want at your dinner party.

A Modest Post on the Nature of Consciousness, Self-Awareness, and What Makes Humans Human

The political campaigns are so dismal that I can’t bring myself to think much about them. So, now for something completely different: Let’s discuss some speculative neurology.

Let me see if I can get enough of this into a small enough space that I don’t get bored with it and it’s semi-intelligible. How brains compute is a moderate obsession of mine, and I’ve come to hold an idea of what I think consciousness and self awareness are. I’ll also take a crack at what makes humans different from other animals.

To start with, remember that evolution hardly ever throws anything away. If something works and provides a survival advantage, it gets conserved. Other stuff comes along later and may leverage the older stuff, or it may simply evolve to accommodate the older stuff.

All life needs to sense its environment, so that it knows how to react to consume resources, and how to avoid being consumed as resources—food. The first step to this involved chemical receptors arrayed on the outside of cells. Some interesting molecule would float by, the receptor would glom onto it, and a cascade of chemical reactions would cause the cell to do something. The “something” could be a simple as opening channel in the cell membrane to consume the molecule, or as complex as using a flagellum to turn toward or run away from the molecule.

Once you get to multi-cellular creatures, more information can be processed if cells specialize to detect certain types of information and relay that information to nerve cells. Nerve cells can then detect patterns of different kinds of information and respond to them with certain actions. Organisms are now capable of learning, but the learned behavior certainly isn’t conscious. It’s merely, “Detect stimulus A, respond with B. Detect stimulus C, respond with D,” and so on.

Animals can go a long way with this type of system, but it becomes energetically inefficient to strew neural processing nets all over the organism. The next step in complexity is to route all stimuli to a central spot, then have that central spot respond with a set of actions that get routed back out to the portions of the animal that can take action. Welcome to the age of brains. A slight step up from this further simplifies the transport of neural signals via a bundle of neurons that feed into the brain, and we’re now into the chordata phyllum.

Note that, evolutionarily, everything makes sense so far. Differentiating from a single-celled organism with receptors to a multi-celled animal with sets of cells doing the receiving and acting to other cells coordinating the receiving and acting starts out as separate organisms in symbiosis, then slowly evolves into true multi-cellularity. Everything is gradual. Similarly, once you get neural nets distributed around the animal’s body, evolution can gradually make things more energy-efficient by having the processing cells migrate together and more densely interconnect.

But brains at this point are still pretty much just bags of loosely-coupled stimulus-response loops. You get lots of variations of neurons that do very specialized tasks, depending on what kinds of stimuli they’re processing, or what kinds of actions—including motor actions—they’re controlling. If you look at the human brainstems, and those of more primitive chordates, you’ll find lots of neural nets like this. There are lots of slightly different tissues performing slightly different actions.

We’re now getting to the point where the array of stimuli from the environment is so large, so varied, that the brain is unlikely to be able to respond to everything that’s going on all at once. As brains have evolved, the environments into which they’ve evolved have become more and more complex. A single-celled animal in a tide-pool has a very simple environment: It “smells” food and tries to move toward it, and maybe it knows how to secrete something nasty or move away to avoid being eaten. An insect or a chordate on land has a bewildering array of features to its environment. Vision is almost essential, and the arms race between predators and prey floods the brain with conflicting stimuli.

Animals at this point have “memory”; they’ve learned to respond to patterns of stimuli with other patterns of response. But they now need a way to prioritize responses. To do this, the brain links lots of disparate neural nets together with another kind of memory, which can weigh conflicting stimuli, hold them together for long enough to select the important ones, and then coordinate a set of behaviors in response. We call this behavior “attention”, and the anatomical structures that contains the working memory for directing attention are the limbic system and the cortex.

Even with specialized tissues, it’s unlikely that all of the neurons are set with response patterns solely from genetic instructions. Things at least need to be fine-tuned, and the animal needs to be plastic enough to recover from damage if it’s to live long enough to reproduce. Brainstem elements therefore “learn”—they modify neural connections to get optimal response from the animal, even though the animal’s sensory and motor systems may change (mostly for the worse) over time.

I don’t know a lot about how those systems learn in things without a cortex. I suspect that learning strategies for various brainstems tissues co-evolved with the tissues themselves. But when an animal has a cortex, it can employ a much more powerful learning strategy.

Part of the attentional system works by detecting a moderate stimulus and, if attention is granted, greatly strengthening it by having the attentional system add further stimulus. This mechanism has another purpose. The combination of attention and stimulus can greatly enhance learning. The attention system can now simulate (no ‘t’) hunks of stimulus so that the motor system can “practice” its response—without actually having to experience the stimulus from the environment or execute the response. We’re pretty sure that this is what dreaming is for. If this is true, then most animals with a cortex should dream.

Being a computer geek, I’ll now make a computer analogy that is only slightly accurate, but will give us a framework for moving forward: The brainstem is the brain’s I/O system, the limbic system and cortex is the CPU, and attention is the active task. The learning mechanism doesn’t have a direct analogy here, but you might think of it as the memory refresh logic, constantly reading in memory and rewriting it. Alternatively, learning can be viewed as subtasks that are launched from an executive in the attentional system.

The more cortex you have, the more sophisticated the relationships are that can be held between the heavily-processed I/O from the peripheral sensory systems and the behaviors to be sent out to the motor (and endocrine) systems. Attention is merely the act of progressively abstracting these relationships and deciding to activate a small number of them instead of activating thousands of them.

I think you can have an organism with attention that isn’t what we would consider self-aware, but I’m willing to bet that the next step is where an organism becomes self-aware. That step is easy to evolve: We already have an attentional mechanism that gets stimulated by the outside world. Those stimuli get more and more hierarchically abstracted, so that a small stimulus pattern to the attentional mechanism can represent an incredibly complex stimulus. We have a word for these stimuli: we call them “concepts”.

We also have the ability for the cortical learning mechanism, which must be closely related to the attentional system, to activate concepts during dreaming. It’s only a small step from dreaming to allowing the attentional system to stimulate concepts during when it’s awake.

This sounds an awful lot like “thinking” to me, and my guess is that the act of thinking is the hallmark of self-awareness. When the attentional system isn’t being consumed with environmental stimuli, it simply activates concepts based on recent experience. The activation of those concepts cause other concepts to be activated, and the attentional system then chooses which concepts to attend to, in the process stimulating those concepts more strongly, which activates still other concepts, and so on.

Note that all of these mechanisms are merely an outgrowth of the more primitive attentional system and the cortical learning system. Again, evolution doesn’t have to add a lot to get greatly enhanced behavior.

If this is right, then almost everything with a cortex is to some extent “self-aware”. So what distinguishes humans from other chordates?

One answer might be “Nothing”. Perhaps human self awareness and processing is merely better as a matter of degree rather than kind. But there is one last piece to the puzzle that it appears that few animals have: language.

At one level, language is easy to extrapolate from what we already have. Since the attentional system can internally stimulate concepts, it may be natural to tag concepts with a motor activity corresponding to a sound. We call these sounds “words”.

A lot of animals are limited in the number of different sounds they can make. It would be natural for them to associate words only with the most important concepts, so that their inventory of sounds is used productively. However, the human vocal tract is capable of producing an astonishing number of sounds. Hence, it can associate a unique word with almost any concept. Note that the key evolution here isn’t really neural; the evolution of the vocal tract naturally leads to more and more words.

Eventually, there are enough words that ever more complex concepts can be created by stringing words together. Patterns of words become new concepts, eventually leading to syntax, which is just an attentional set of tricks to allow the sequencing of words.

At the end of this process, we have a whole new emergent property, in which transient concepts can be sent as stimuli from one brain to another, via words and syntax. Some of those concepts will become new, permanent concepts. Others may persist just long enough to allow groups to coordinate actions, socialize, or merely have a nice chat about last night’s basketball game. But the end result is a whole new level of emergent behavior that only humans (and perhaps a couple of other species) possess.

Those last tiny steps of evolution are interesting. If I’m right, being human is possible only because we can make a lot of distinct noises, which has led to an explosion in the number of concepts we can handle, which has further driven the expansion of the cortex. If you could give other animals the ability to make more noises (say, by grafting vocoders into their brains at or near birth), you might discover that the differences between humans and animals really was more one of degree than of kind. Perhaps words and syntax are merely an outgrowth of the need to handle more and more concepts. That might make for a future where the definition of humanity might have to be broadened considerably.

UPDATE 9/19/12:  Just fixed some formatting.