General Intelligence and Context Switching

General intelligence is a remarkable adaptation. We humans are able to learn a near infinite variety of things, from language to physics to etiquette to chess. Our brains never “fill up”; in fact, increased capability in one domain often enhances our ability to learn others. It doesn’t take much for us to learn, with facts often committed to memory with a single observation, and improvements in more complex domains such as chess occurring in the span of hours. It seems one of the keys to our ability lies in context switching – we’re able to quickly understand the context of a given situation, and then leverage all the strategies we’ve learned for success in that context. For example, when presented with physics problems, formulas such as “Force = Mass * Acceleration” and “Position = ½ * Acceleration * Time²“ naturally come to mind, but switch out the physics problems for a chessboard and we instantaneously adjust, bringing to bear instead all the pattern recognition and opening memorization which drives success in chess (and setting aside the physics formulas). Our lives are a series of different contexts, and our brains are extremely adept at figuring out our current context and putting the right set of skills forward.

Scientists in different domains have examined this idea, with each providing a different visual for this context switching. Marvin Minksy (a giant of computer science and AI) viewed the mind as composed of many agents in a “Society of Mind”:

I’ll call “Society of Mind” this scheme in which each mind is made of many smaller processes. These we’ll call agents. Each mental agent by itself can only do some simple thing that needs no mind or thought at all. Yet when we join these agents in societies – in certain very special ways – this leads to true intelligence.

(Minsky, Society of Mind)

The neuroscientist Christof Koch (president of the Allen Institute for Brain Science) views the mind as composed of “zombie agents”:

Neurologic and psychological sleuthing has uncovered a menagerie of specialized sensory-motor processes. Hitched to sensors – eyes, ears, the equilibrium organ – these servomechanisms control the eyes, neck, trunk, arms, hands, fingers, legs, and feet, and subserve shaving, showering, and getting dressed in the morning; driving to work, typing on a computer keyboard, and text messaging on your phone; playing a basketball game; washing dishes in the evening; and on and on. Francis Crick and I called these unconscious mechanisms zombie agents. Collectively, this zombie army manages the fluid and rapid interplay of muscles and nerves that is at the heart of all skills and that makes up a lived life.

(Koch, Consciousness: Confessions of a Romantic Reductionist)

The psychologists Ecker, Hulley, and Ticic, in their book Unlocking the Emotional Brain, see the mind as made up of schemas, or individual world models (their focus is on the role of emotion in shaping these models):

Unlocking the Emotional Brain’s premise is that much if not most of our behavior is driven by emotional learning. Intense emotions generate unconscious predictive models of how the world functions and what caused those emotions to occur. The brain then uses those models to guide our future behavior. Emotional issues and seemingly irrational behaviors are generated from implicit world-models (schemas) which have been formed in response to various external challenges. Each schema contains memories relating to times when the challenge has been encountered and mental structures describing both the problem and a solution to it.

(Summary by Kaj Sotala, originally read here)

Whether we view the phenomenon as consisting of agents, zombies, or schemas, it’s clear that there’s a shared thread among all these views, one with numerous related yet distinct “operational strategies” (I’ll refer to them in this way through the remainder of the post) coming together to form general intelligence. We have all these strategies inside of us – or really, all these strategies are us. When we learn a new skill – chess, for example – we form a new strategy, one dedicated solely to learning the patterns of the board and memorizing openings. Interestingly, this chess strategy is itself made up of many semi-distinct strategies, for tasks like learning board patterns and memorizing openings are best accomplished separately. We can see a hierarchy forming – beneath our umbrella “I”, we have an operational strategy dedicated to game-playing (bringing with it competitive spirit, focus, etc.), and within that strategy we have another dedicated to chess (bringing with it general chess understanding and goals), and continuing down the hierarchy we see additional strategies dedicated to particular aspects of chess (such as tactics, openings, and the endgame). 

In practice, the lines between these different parts of the hierarchy are more blurred, but the overall idea still stands. Each operational strategy is designed for a particular context, and the core driver of our general intelligence is the fact that we can effortlessly switch between these strategies as contexts require.

Looking at general intelligence in this way raises at least two major questions:

1. How does our brain “know” which context we’re in / which operational strategy to apply?
2. How does our brain “know” when to branch off and form a new operational strategy, vs. simply updating an existing strategy? For example, why do we form a new strategy for chess instead of updating our checkers strategy to account for chess moves?

For question 1, we can gain some insight by looking at a much simpler brain, that of the lamprey:

How does the lamprey decide what to do? Within the lamprey basal ganglia lies a key structure called the striatum, which is the portion of the basal ganglia that receives most of the incoming signals from other parts of the brain. The striatum receives “bids” from other brain regions, each of which represents a specific action. A little piece of the lamprey’s brain is whispering “mate” to the striatum, while another piece is shouting “flee the predator” and so on. It would be a very bad idea for these movements to occur simultaneously – because a lamprey can’t do all of them at the same time – so to prevent simultaneous activation of many different movements, all these regions are held in check by powerful inhibitory connections from the basal ganglia. This means that the basal ganglia keep all behaviors in “off” mode by default. Only once a specific action’s bid has been selected do the basal ganglia turn off this inhibitory control, allowing the behavior to occur. You can think of the basal ganglia as a bouncer that chooses which behavior gets access to the muscles and turns away the rest. This fulfills the first key property of a selector: it must be able to pick one option and allow it access to the muscles. Many of these action bids originate from a region of the lamprey brain called the pallium [the region which evolved into our cerebral cortex]…

Each little region of the pallium is responsible for a particular behavior, such as tracking prey, suctioning onto a rock, or fleeing predators. These regions are thought to have two basic functions. The first is to execute the behavior in which it specializes, once it has received permission from the basal ganglia. For example, the “track prey” region activates downstream pathways that contract the lamprey’s muscles in a pattern that causes the animal to track its prey. The second basic function of these regions is to collect relevant information about the lamprey’s surroundings and internal state, which determines how strong a bid it will put in to the striatum. For example, if there’s a predator nearby, the “flee predator” region will put in a very strong bid to the striatum, while the “build a nest” bid will be weak…

Each little region of the pallium is attempting to execute its specific behavior and competing against all other regions that are incompatible with it. The strength of each bid represents how valuable that specific behavior appears to the organism at that particular moment, and the striatum’s job is simple: select the strongest bid. This fulfills the second key property of a selector – that it must be able to choose the best option for a given situation…”

With all this in mind, it’s helpful to think of each individual region of the lamprey pallium as an option generator that’s responsible for a specific behavior. Each option generator is constantly competing with all other incompatible option generators for access to the muscles, and the option generator with the strongest bid at any particular moment wins the competition.

(excerpt from The Hungry Brain by Guyenet, taken from this post)

While our behaviors are far more complex, we can see from the lamprey how our brain might go about selecting an operational strategy. The work occurs in parallel, with each operational strategy (instantiated in various parts of the cortex) “evaluating” its applicability, and passing on that value to the selector (the striatum / basal ganglia), which picks one. This setup adds another requirement to our operational strategies – not only must they “know” what to do, they also need a way of evaluating whether to do it. In reality, it seems there’s room for this whether evaluation to be more distributed, with some part happening in the basal ganglia (perhaps conversion of different bids into the same “units”, or something comparable), as complete evaluation would seem a tall order for an individual strategy (evaluation is complex, and at the lowest levels of the hierarchy these strategies are comparatively simple). 

While the first question can be answered in broad strokes at the component level, without a detailed understanding of the algorithms of the brain, the second question (how does our brain “know” when to branch off a new operational strategy) does not yield as easily to this level of analysis. While we know that some determinants must inform the branching strategy (for example, the level of similarity of the current context to that addressed by other existing operational strategies), this is not a satisfactory answer; what we really want to know is exactly how these determinants function. What happens in our brain when we start playing chess for the first time? We can imagine circuits being activated in novel combinations, but novel combinations are nothing new – actions as simple as reading a new book or traveling to a new place drive similar degrees of newness. How do our brains “know” that the chess newness requires developing a distinct strategy, while new books and places can generally (but not always!) be addressed using existing strategies? This appears to be a hard problem, and one which will be difficult to answer without a better understanding of how the brain actually works.

This problem also has direct ramifications for the field of AI, as until we make more progress here, it seems likely that we’ll be unable to advance beyond the “narrow” models that exist today. Most of today’s systems are constructed with a particular context in mind, for which a single operational strategy is required (note that “single” here is a loose definition, and depends on the level of the hierarchy at which you look – for as we saw earlier chess has a number of sub-strategies below it). We train a model to recognize images, and that’s all the model is asked to do; no context switching is required. We’ve gotten quite adept at creating models which solve these narrow problems effectively, even in complex domains like image recognition and video game playing. However, most useful work performed today by humans requires a large degree of context switching: a cashier must quickly shift from handling bills to answering a customer’s question; a janitor must transition from putting out the trash to fixing the recently broken window; an office worker must be able to follow up crunching numbers in Excel with a conversation on the significance of the data. Though we could likely construct models which would handle each of these tasks effectively individually, we haven’t made much progress on doing them all (together with the nearly infinite variety of other tasks required) effectively. Making progress here will require a better understanding of how to build context switching into our systems, together with improved abilities to adapt to new contexts and form new operational strategies. It seems our best path forward will be to try and understand how our brains handle such a wide variety of contexts so seamlessly (including, luckily for us, the context of neuroscience!)