Modeling Control Without Controlling the Model

The great hope of Henry Markram’s Blue Brain project (recently discussed here and here) is that it will demonstrate both that consciousness and agency are emergent properties of an entirely mechanistic system like the brain and how that could possibly be true.  Despite Markram’s headline-grabbing claim at TED last week that he’s 10 years away from a complete silicon brain, that knowledge, perhaps the holy grail of neuroscience, still lies rather far from our grasp (although it’s probably fair to say that we can now see it).

What we have today, however, is knowledge of increasingly complex individual components of the complete physical explanation for the human experience.  To stretch my Apollo metaphor from last week perhaps near its breaking point, it might be fair to say that the field is currently in the “Gemini” stage.  Each Gemini mission demonstrated that individual components of a lunar mission–extra-vehicular activity, orbital rendezvous, docking–were even possible before they could all be woven into the fabric of a full-fledged moonshot.

Likewise, contemporary investigations at the cellular level of the brain attempt to understand core processes in isolation.  I mentioned a critical example in my post about the Blue Brain project last week, and I felt like I should giv e it more attention because it’s an example of an important step towards the ultimate goal (also, because I finally got around to reading the paper).  The study in question was published earlier this month in The Journal of Neuroscience. Authored by Chung-Chuan Lo, it’s the result of a collaboration between the labs of Jeffery Schall at Vanderbilt and Xiao-Jing Wang at Yale.

Schall’s lab has long been involved in understanding the neural basis of cognitive control.  They study the activity of individual neurons in the brains of rhesus macaques as the monkeys perform a seemingly simple task.  The animals are trained to stare at a central point until it disappears, replaced by a new point in the periphery of the animal’s vision.  To get a reward, the monkey just has to look at the new dot.  There’s a catch, though.  Some of the time, the central dot will reappear.  On those trials, the monkey will be rewarded only if it restrains itself from looking to the side.

So the task seems simple, but the processes it recruits are central to our ability to perform goal-directed or intentional actions.  After-all, if we couldn’t inhibit strong habitual responses in the service of an abstract goal, we’d be nothing more than zombies or automatons.  But any responsible neuroscientist could tell you that it’s wrong to think of habits or instincts encoded in our brains as being controlled by “our minds,” as if the two are separable.¹  Which is not to say that it’s easy to understand how they are the same thing.

Schall’s research has identified several classes of neurons thought to be involved.  Some fire when the middle fixation point is present and cease activity when it vanishes.  Others demonstrate a steady increase in activity between that time and when the monkey looks towards the second point.  On trials when the middle point reappears, the first type of neurons rapidly start firing again.  When the monkey successfully restrains itself, the second type of neurons quickly calm down.  Schall has proposed that the decision whether to move is the result of a competition between two independent processes, a GO process and a STOP process, that are apparent in this activity.

The obvious limitation to this approach is that it looks at single cells in isolation.  Of course, there are thousands of neurons involved in this behavior, and the relevant interactions between those neurons can only be inferred from the individual data.  This is where Wang’s work comes in.  Wang’s lab focuses on biophysically realistic modeling of neural circuits.  His models are considerably simpler than the Blue Brain, but the dynamics of each individual “neuron” are carefully tuned to mimic their real-life counterpart.

Wang’s model for this experiment consisted of a handful of discrete groups of several hundred simulated neurons in the circuit.  As suggested by the neurophysiological research, one group responds to the central stimulus and restrains movement, while another group responds to the peripheral stimulus and initiates it.  The first group are further influenced by a “top-down” signal that modulated the level of control in the circuit.  This signal likely arises from the prefrontal cortex, and it probably demonstrates the effects of conscious attention-focusing.

The neural network model, when properly tuned, exactly duplicated the behavioral results of actual monkeys performing the task.  This makes a strong case, in other words, that attempts at self-control result in a competition between independent processes and that the amount of top-down control biases the competition appropriately.

What’s really interesting, though, is looking at the circuit as a whole.  Because of the nature of the model, it acts like a dynamic system that exhibits attractor dynamics around one of the two outcomes.  That’s a way of saying that it produces large-scale coordinated across the hundreds of simulated neurons that was never intentionally programmed into the model.  Each “neuron” was tuned individually, but the way they act together to produce a movement emerges naturally.  It’s likely that this is what’s going on in the Blue Brain video when you see spontaneous waves of activity sweep across the simulated neocortical column after only a few silicon cells are activated.

So this paper models control without actually controlling the model.  The Wang model shows that a mechanistic system, based pretty closely in the actual brain, can control itself, at least in response to an external stimulus.  The next step will be to show that such a system can control itself based on completely internal guidance.  And, in doing so, we’ll step closer and closer to understanding the whole brain and, thus, ourselves.

Lo, C., Boucher, L., Pare, M., Schall, J., & Wang, X. (2009). Proactive Inhibitory Control and Attractor Dynamics in Countermanding Action: A Spiking Neural Circuit Model Journal of Neuroscience, 29 (28), 9059-9071 DOI: 10.1523/JNEUROSCI.6164-08.2009

¹ Of course, there are some neuroscientists who would tell you the mind and the brain are separable, and it’s not as if they’re irresponsible or shoddy scientists, but I do happen to think that they’re wrong.