By now I've written a bunch of blog posts on brain architecture and algorithms, not in any particular order and generally interspersed with long digressions into Artificial General Intelligence. Here I want to summarize the key ideas in one place, to create a slightly better entry point, and something I can refer back to in certain future posts that I'm planning.

I am trying to paint a picture here, but this post won't really justify it, let alone prove it. Personally, my main justification is: I think I can basically understand how every piece can be implemented with biologically-plausible components, and I think I can basically understand how the pieces work together to explain absolutely everything in human cognition and psychology. (Try me!) Needless to say, I could be wrong in both the big picture and the details.

Pretty much everything I say here and its opposite can be found in the cognitive neuroscience literature. (It's a controversial field!) I make no pretense to originality but can't be bothered to put in actual references. Ask me if you're interested.

Let's start with a diagram, similar to the one I put in Inner Alignment in the Brain:

So let's start in:

Two subsystems: "Neocortex" (including thalamus, hippocampus, basal ganglia), and "Subcortex"

The neocortex is the home of "human intelligence" as we would recognize it—our beliefs, goals, ability to plan and learn and understand, every aspect of our conscious awareness, etc. etc. (All mammals have a neocortex; birds and lizards have an analogous structure called the "pallium".) Some other parts of the brain (thalamus, hippocampus, basal ganglia) help the neocortex do its calculations, and I lump them into the neocortex subsystem. I'll use the term subcortex for the rest of the brain (midbrain, amygdala, etc.).

Cortical uniformity

The neocortex is, to a first approximation, architecturally uniform, i.e. all parts of it are running the same generic learning algorithm in a highly-parallelized way. The caveats (spelled out in more detail at the link) are:

There are sorta "hyperparameters" on the generic learning algorithm which are set differently in different parts of the neocortex—for example, different regions have different densities of each neuron type, different thresholds for making new connections (which also depend on age), etc.
There is a gross wiring diagram hardcoded in the genome—i.e., set of connections between different neocortical regions and each other, and other parts of the brain. These connections later get refined and edited during learning. These speed the learning process by bringing together information streams with learnable relationships—for example the wiring diagram seeds strong connections between toe-related motor output areas and toe-related proprioceptive (body position sense) input areas.

What is the neocortical algorithm?

"Analysis by synthesis" + "Planning by probabilistic inference": "Analysis by synthesis" means that the neocortex searches through a space of generative models for a model that predicts its upcoming inputs (both external inputs, like vision, and internal inputs, like proprioception and reward). "Planning by probabilistic inference" (term from here) means that we treat our own actions as probabilistic variables to be modeled, just like everything else. In other words, the neocortex's output lines (motor outputs, hormone outputs, etc.) are the same type of signal as any generative model prediction, and processed in the same way........... g As shown in this figure, the neocortex builds a world-model in the form of collection ("society") of generative models, and it continually searches through this collection for models that make correct predictions, and favoring those models which predict future reward. in other words, you are treating your own future actions as probabilistic variables to be modeled just like any other part of the universe. We take adaptive actions only because, as mentioned above, we favor generative models that predict we will get rewarded. (I take the term "planning by probabilistic inference" from this non-brain-related paper.)

Planning by probabilistic inference: In line with mainstream predictive coding, I claim that the neocortex's outputs are the same type of signal as a generative model prediction—
Aside: Is this the same as predictive coding / free-energy principal? My account differs from mainstream predictive coding (Karl Friston, Andy Clark, etc.), by (1) emphasizing multiple competing generative models rather than "a prior", and (2) restricting discussion to the neocortex—I don't think predictive coding is a helpful way to think about other parts of the brain. For both of these, I'm not sure to what extent this is a fundamental difference versus a difference of emphasis & intuitions.
Compositional generative models: Each of the generative models consists of predictions that other generative models are on or off, and/or predictions that input channels (coming from outside the neocortex—vision, hunger, reward, etc.) are on or off. ("It's symbols all the way down.") All the predictions are attached to confidence values, and both the predictions and confidence values are, in general, functions of time (or of other parameters—I'm glossing over some details). The generative models are compositional, because if two of them make disjoint and/or consistent predictions, you can create a new model that simply predicts that both of those two component models are active simultaneously. For example, we can snap together a "purple" generative model and a "jar" generative model to get a "purple jar" generative model. They are also compositional over time (you can make a generative model that says "generative model X happens and then generative model Y happens").
PGM-type message-passing: Among other things, the search process for the best set of active generative model involves something at least vaguely analogous to message-passing (belief propagation) in a probabilistic graphical model. Dileep George's vision model is a well-fleshed-out example.
Hierarchies, non-hierarchies, feedforward, and feedback signals:
- Hierarchies are a special case of compositional generative models. A generative model for an image of "8" makes strong predictions that two "circle" generative models are positioned on top of each other, and the "circle" generative model in turn makes strong predictions that certain contours and textures are present in the visual input.
- However, not all relations are hierarchical. The "is-a-bird" model makes a medium-strength prediction of the "is-flying" model, and "is-flying" makes a medium-strength prediction of "bird". Neither is hierarchically above the other.
- ...
- (There's no good terminology here. If I say "feedforward" and "feedback", it's intuitively sensible for sensory inputs (where information moves "forward" from the input channels), but feels backwards for motor outputs. If I say "bottom-up" and "top-down", which are classed as "feedback" even though it feels like the forward direction), but feels backwards for motor outputs ( a bit confusing because the top-down
- The two types of signals in the neocortex are "top-down" and "bottom-up" signals (a.k.a. "feedback" and "feedforward" respectively, but I dislike that terminology because it works for sensory inputs but it kinda feels backwards for motor outputs).
- Top-down signals encode predictions for bottom-up signals. The predictions have confidence levels, spanning the range from "If I had to guess, I suppose I'd say that this signal will be active" to "This signal will be active or I will eat my hat!!!" When many high-confidence predictions are in error, a bottom-up signal goes into the next rung of the hierarchy, eventually bubbling up to conscious attention.
- Input data coming into the neocortex from outside is a special case of "bottom-up". Output signals exiting the neocortex are a special case of "top-down".
- ..................... of which include types As described here, the neocortex's algorithm involves searching through a space of compositional generative models for a model that matches the current What is the neocortex's algorithm? it basically builds a bunch of predictive models of its inputs, outputs, and rewards, with a bias towards models that have made correct predictions in the past (self-supervised learning), that are compatible with other models (Bayesian priors, sorta), and which predict that high reward will happen (reinforcement learning). The main types of signals in the brain are called "top-down" [TODO]
- Note that my version of predictive coding is narrower than the story you get from Karl Friston, Andy Clark, etc. They conceptualize predictive coding as a theory of the entire brain, whereas I think it's only helpful for describing the neocortex. There are some other differences too.
Blank-slate neocortex (but non-blank-slate subcortex): The neocortex starts out as a "blank slate": Just like an ML model with random weights, it cannot make any correct predictions or do anything useful until it learns to do so from previous inputs, outputs, and rewards. The subcortex, however, is absolutely not a blank slate, and that is where we get all our specific instincts and dispositions (see my general discussion of human instincts here). (By the way, I am not saying that the neocortex's algorithm is similar to today's ML algorithms—there's more than one blank-slate learning algorithm!)
The subcortex tries to steer the neocortex towards biologically-adaptive behaviors: The blank-slate neocortex can learn to predict input patterns, but it needs guidance to do biologically adaptive things. So one of the jobs of the subcortex is to try to "steer" the neocortex, and the subcortex's main tool for this task is its ability to send rewards to the neocortex at the appropriate times. Everything that humans reliably and adaptively do with their intelligence, from liking food to making friends, depends on the various reward-determining calculations hardwired into the subcortex.
The subcortex is relatively stupid—from its perspective, the neocortex is a giant black-box machine doing utterly inscrutable calculations. So how does the subcortex figure out when to send rewards to the neocortex? Two ways:
1. The subcortex can learn what's going on in the world via its own, parallel, sensory-processing system. Thus, for example, we have the well-known visual processing system in our visual cortex, and we have the lesser-known visual processing system in our midbrain (superior colliculus). Ditto for touch, smell, proprioception, nociception, etc. While they have similar inputs, these two sensory processing systems could not be more different. The neocortex fits its inputs into a huge, open-ended predictive world-model, but the subcortex instead has a small and hardwired "ontology" consisting of evolutionarily-relevant inputs that it can recognize like faces, human speech sounds, spiders, snakes, looking down from a great height, various tastes and smells, stimuli that call for flinching, stimuli that one should orient towards, etc. etc., and these hardwired recognition circuits are connected to hardwired responses. For example, babies learn to recognize faces quickly and reliably in part because the midbrain sensory processing system knows what a face looks like, and when it sees one, it will saccade to it, and thus the neocortex will spend disproportionate time building predictive models of faces.
2. The subcortex can see the neocortex's outputs, which (as in the diagram above) also includes predictions, imagination, memories, and empathetic simulations of other people. For example, if the neocortex never predicts any reward, then the subcortex can guess that the neocortex has a grim assessment of the future—I'll discuss that particular example much more in an upcoming post.
  - To squeeze more information out of the neocortex, the subcortex can also "teach" the neocortex to reveal when it is thinking of one of the situations in the subcortex's small hardwired ontology. For example, if the subcortex rewards the neocortex for cringing in advance of pain, then the neocortex will eventually start sending weak cringe-motor-commands when imagining future pain, or when empathically simulating someone in pain ... And the subcortex can detect that, and have specific hardwired responses. (There may be other possible mechanisms too, that do the same thing.)
  - (See Inner Alignment in the Brain for more on both of these.)
s

All of these are slight oversimplifications, and I won't justify here why I believe all these things, except to say that I'm really loving this framework so far. I feel like I can relate it to both lower levels (i.e. how each component is implemented) and higher levels (i.e. how human behavior comes out of it), and everything fits together beautifully.

Neuroscience