Defrosting our Views on Intelligence
Introduction
Succeeding as a data scientist in 2020 depends on a well-tuned BS meter. Machine Learning (ML) and Artificial Intelligence (AI) are hot topics — they promise immense potential value while being cool — meanwhile, everyone (including you and me!) has a poor grasp of the AI/ML field. When you mix greed and rampant Illusory Superiority (Dunning-Kruger), you end up with a thriving population of shysters and imposters (Meanwhile, inducing impostor syndrome in many of the real, humble practitioners).
Mostly, I chalk this up to the rapid growth of the field. However, it also partly stems from a lack of real authority on the fundamentals of intelligence: there’s no academic consensus or broad institutional understanding of what we mean by “intelligence”, “learning”, or “abstraction”. As such, it becomes very difficult to meaningfully discuss the big picture around AI. Which is rather annoying, because it’s something we both want and need to discuss!
So here begins my assessment of “intelligence” and what it means to us. I will mostly use the term “cognition” going forward, to indicate the specific processes and abilities that humans possess.
Caveat Lector: I studied physics, and claim no professional expertise in understanding intelligence. This is just my hot take. Consume this like the scientific equivalent of fast food: be skeptical of the ingredients and enjoy what you can!
Shaping Up
Imagine that human academic disciplines take on some physical state. This state reflects the stability, predictive ability, and integrated knowledge a field has. Our theories on electromagnetism, for example, might look like a single crystal — solid, ordered, and connected — as they are able to correctly predict and explain our everyday encounters with light and electricity. In fact, Richard Feynman described quantum electrodynamics as “the jewel of physics”, so I’ll give him the prior art nod. In other words, this subfield of physics is “solved” to a meaningful degree — mostly because physics is the easiest science on which to make this type of progress.
On the other end of the spectrum, we have fields that are quite difficult to tame — namely, anything involving interacting humans, like markets and political movements. We might come up with some explanations, handy rules, and descriptive equations, but as they aren’t directly grounded in the basic sciences it is difficult to develop quantitative theories and apply predictive rigor. Take economics, where we can make some use of notions of supply and demand, but we don’t have a full theory that can predict the price of any good. These disciplines might be thought of as a type of soft clay — moldable into different large shapes, but unfixed and lacking smaller structure.
Studying Cognition
…the field as a whole seems to have lost impetus, focus, and recognition.*
Cognitive science, a young inter-disciplinary field representing our best attempt to combine our knowledge and understanding of how the brain works, is in a different state. The above quote comes from an *article (pdf here) from summer 2019 which bemoans the failure of the field to coalesce into a mature state. In their analysis, traditional psychology has dominated the field, to the detriment of more hardened fields like neuroscience and AI, and they point to a lack of curricular consensus — a student would learn different things from different universities. You have a collections of contributions from some noteworthy folk, like Douglas Hofstadter, Marvin Minsky, and Daniel Dennet; but these don’t link up into anything coherent. I imagine the field manifesting as a layer of small crystals, as you’d find in a freezer in need of a cleaning. There’s lots of little quantitative facts, but it’s not clear if they’re going to work together and produce anything meaningful. Time to defrost it! Here goes my attempt to make some basic sense of it…
Applying the Heat
The ultimate goal I have is to develop a working sketch of a theory of cognition: the first layer of fundamental abstraction of what’s happening in our heads. So, after clearing the slate of noise, what guiding principles can we use to set our path towards this sketch? I came up with these motivating questions:
- What differentiates us from other animals?
- What is the simplest explanation of this difference?
- Does this fit with what we perceive in ourselves?
A tall order, so we definitely need to bring as much firepower as we can, and apply a divide and conquer strategy. First off, we’ll make a few necessary assumptions to avoid getting hung up on details:
- Naturalism: no supernatural forces are at work.
- Universality: we share a common, fundamental human condition
- Reducibility: mental processes are reducible to smaller/simpler things
- At least two types of thought process exist: fast and slow
The first point should be self-explanatory. By the second, I want to preclude any assumption that humans can be divided into groups that “think differently”. In other words, I’m assuming the simplest case: we are, by and large, similar in our basic function. Our brains have the same regions and these tend to wire up in a certain way, at some large scale.
Reducibility addresses the differences we do see, and says they arise from combining the building blocks of the (as opposed to some emergent, irreducible complexity). Humans exhibit memory, learning, emotions, pattern-recognition, etc and we can abstract cognition, and its surrounding context, into these pieces.
In the last point, I build on the classification popularized by Kahneman, where he suggests we have two systems of thought: one fast, reflexive, unconscious; the other slow, deliberative, conscious. I figure most people can accept these results (they match everyday experiences) and it gives us a good starting point for classifying different types of mental processes.
A Map of Thoughtland
The rest of this article is devoted to scrutinizing this picture: my dinner-napkin-level drawing of what might reasonably happen in our brains. I’ve concocted this as a handy and suggestive reference, nothing more. The rows represent three states of information flow, while the columns are a hierarchy of function. Inputs and Outputs should be self-explanatory, while for Internal states, I am referring to something the brain has learned to do — some kind of adjustable processing. Passive and Reactive functions are both part of fast, non-deliberative thinking (System 1): passive further indicates continuous, common functions while reactive processes are one-off, notable events. Active functions are our slow, deliberative thinking processes, of which we are continually aware.
Let’s imagine how our body could determine how to sweat. Abstractly, all that’s required is a measure of our body temperature (T) and a response based on how far that temperature lies above a learned threshold (T_max). So we can imagine T as a passive input, while T_max is some hard-coded, genetic constant. The internal process is then a function taking T, T_max and outputting a signal to sweat at an appropriate rate.
The arrows represent the possible interactions between these areas, and contain more of the interesting bits, I think. The downward areas are more straightforward, representing the passing of inputs to, and outputs from, the internal functions. The two diagonal arrows, up and to the right, indicate that lower level abstractions are what generate the inputs to the higher level processes.
Now the leftward arrows are, I think, interesting. I imagine these as a slow process that is training our internal functions. Consider the following: most of us don’t have an appropriate sense of danger at birth — we’re at the mercy of hot stoves, dangerous animals, etc. However, we can learn to react to these situations in an unconscious way, through some type of reinforcement. When we see the glowing red iron or shark fin, we now immediately recognize the danger and can take quick action. Similarly, the emotions we feel towards people, objects, and concepts are mostly learned; the feelings you have towards your friends was very different when you first met them. This is guided by how we consciously think about what’s happening, what they say and do, whether we should trust them.
And now time for the punch line:
The answer to humanity, intelligence, and everything is in that one semicircular arrow that feeds our internal, active state back into itself.
That arrow is where humans really differentiate themselves from other animals, where “intelligence” lies (hold on for my new definition), and where our sense of free will arises. At least, according to the 2020 version of myself. It’s the process for building upon our conscious knowledge, which lets us construct fictions and abstract concepts. If it helps, you could imagine a 4th column and beyond to the right, where these abstractions live at higher levels. If you go far enough, you might reach where M.C. Escher spent most of his time.
Are we really the only animal that has this feedback loop? Read further…