Chapter 11: Generative Models and the Neocortical Mystery

Core Idea

The neocortex seems to do many different things (vision, hearing, touch, movement, language), yet it looks the same everywhere. Mountcastle proposed that it is one repeated microcircuit doing one computation on different inputs. Bennett argues that this computation is a generative model. The neocortex renders an inner simulation of the world, matches it to the senses (perception), and can run it with no input at all (imagination, dreaming). Its real power, and the “world model” that AI still lacks, is that this simulation can be explored on its own.

Frameworks Introduced

• Mountcastle’s crazy idea (the neocortical column):
  • The three observations: neurons within a vertical column (about five hundred microns across) respond alike, while neighboring columns respond differently; connections run mostly within a column rather than between columns; and the neocortex looks almost identical everywhere and across mammals (rat, monkey, human).
  • The conclusion: the neocortex is a sheet of duplicated microcircuits. Every column does the same thing, and regions differ only in their input and where they send their output.
  • Evidence: in 2000, MIT scientists rewired retina input into the auditory cortex of young ferrets, and the ferrets could see fine. The visual cortex of congenitally blind people responds to sound and touch. After a stroke, nearby areas take over the dead area’s function.
  • Why it matters for AI: instead of decoding over ten billion neurons and trillions of connections, perhaps we only need to understand one circuit, repeated a million or so times, with about a million connections.
• The three peculiar properties of perception (the evidence from illusions):
  1. Filling in: the mind adds what is missing. Figure 11.4 shows the word EDITOR drawn with most of its strokes missing, three notched black discs that make you see a white triangle, radiating spikes that make you see a sphere, and two black curved shapes that look like a bar with something wrapped around it.
  2. One at a time: when an image has two readings, you see only one at a time. Figure 11.5 shows the Schroeder staircase (stairs seen from above or from below), the Necker cube (either face can be the front) and the duck-rabbit. The cocktail-party effect is the hearing version.
  3. Can’t unsee: once you get a good interpretation of vague input, you keep it. Figure 11.6 is a black-and-white blob image that suddenly becomes a frog’s face among leaves once you see the gray photo it came from.
  • Helmholtz’s explanation: you perceive what the brain infers is there, which he called inference. “you don’t perceive what you actually see, you perceive a simulated reality that you have inferred from what you see.”
• Generative models: recognizing by simulating (the Helmholtz machine, Hinton and Dayan, 1995):
  • How: a network with forward (upward) and backward (downward) connections. In recognition mode, sensory input flows up to causes at the top, and the backward weights learn to re-create the input. In generative mode, a representation at the top flows down to make sensory input, and the forward weights learn to recognize what was made. Figure 11.7 labels the top layer Causes, the middle Perception / Latent representation, and the bottom Sensory input.
  • Result: trained only on pictures of handwritten digits, with no labels, it learned to recognize digits, to generalize across handwriting, and to generate new digits. Figure 11.8 places a grid of real images of 0–9 beside a nearly identical grid of imagined images that the network generated. “This network has learned to recognize by generating its own data.”
  • Why it works: recognition and generation each train the other, so the network settles without being told the right answer. Failure mode: with no input to constrain it, generation drifts and the model hallucinates.

Key Concepts

• Neocortical column / microcircuit: the repeated unit. It has six layers (earlier cortex had three). Layer 4 gets input from the thalamus, layer 5 projects to the basal ganglia, thalamus and motor areas (figure 11.3 labels this output To thalamus, basal ganglia, & brainstem), and layer 6 projects back to the thalamus. Figure 11.3 also shows thalamic input reaching layer 1.
• Perception by inference: matching an inner simulation to sensory evidence.
• Generative model: a model that learns to recognize by generating its own data and comparing it to real data. Modern examples include StyleGAN2 (thispersondoesnotexist.com), deepfakes, AI art and GPT-3.
• Latent representation: the top-level neurons. Figure 11.10 shows that changing them edits a face, for example beard or no beard, young or old, brown or blond hair, and mouth open or closed.
• Constrained hallucination: what some neuroscientists call normal perception. Without input it becomes unconstrained, as in Charles Bonnet syndrome, where people who lose input from their eyes hallucinate for months.
• Wake-sleep algorithm: Hinton’s name for the Helmholtz machine’s training. Recognition is “awake” and generation is “asleep.”
• World model: an inner simulation that predicts the results of actions and can be searched to plan. Yann LeCun says it is what AI lacks.

Mental Models

• Think of the neocortex as a world-prediction machine. Reflex circuits predict reflexes, and the basal ganglia’s critic predicts reward, but the neocortex predicts all sensory data.
• Think of perception and imagination as “two sides of the same coin.” The same neurons fire when you see a face and when you imagine one, and damage that blinds the left visual field also stops people from imagining things there.
• Think of the neocortex as always asking what 3D world would explain its input. Inferring causes means finding the simulated world that best produces what you sense.
• Think of dreaming as forced generation. Awake, we mostly recognize. Sleep may rebalance the model, which is why sleep deprivation causes hallucinations in mammals and birds.
• Think of the neocortex’s generality as coming from narrow built-in assumptions. It may assume input comes from 3D objects that exist apart from the self and move over time, so it does not have to learn space, time, or self versus other.

Anti-patterns

• Believing the neocortex does many different things. Mountcastle’s view is that it does one thing on different inputs.
• Assuming the brain recognizes objects through supervised learning. Brains are not told the right answer, so they must learn unsupervised.
• Treating imagination as a separate system from perception. In the neocortex they are the same area, and you cannot do both at once. Pupils dilate when people imagine, as the brain stops processing visual data.
• Trying to make AI general by making it less constrained. Bennett suggests this may be backward, because the neocortex may be general precisely because of its narrow assumptions.
• Assuming language and logic are what AI is missing. LeCun argues it is something more primitive: “The ability to learn such world models is what’s missing from AI systems today.”

Rules of Thumb (author’s wording)

• The neocortex “is a world-prediction machine.”
• Generative models “learn to recognize things in the world by generating their own data and comparing the generated data to the actual data.”
• Its power shows most “when the simulation in your neocortex becomes decoupled from the real external world around you.”

The Evolution of Prediction

	Early bilaterians	Early vertebrates	Early mammals
What is predicted	Reflex activation	Future rewards	All sensory data
Where	Reflex circuits	Cortex and basal ganglia	Neocortex

Worked Example

Testing Mountcastle’s idea with rewired ferrets (MIT, 2000).

1. Prediction: if every column runs the same algorithm, the visual and auditory cortex should be interchangeable.
2. In young ferrets, cut the input from the ears and route the retina’s input to the auditory cortex instead.
3. If Mountcastle is wrong, the ferrets end up blind or visually impaired, because the auditory cortex cannot handle visual input.
4. Result: the ferrets saw fine, and the auditory area responded to visual stimuli just as visual cortex does.
5. Conclusion: the computation is general. What a region does depends on its input, so understanding one column could explain the whole neocortex.

A second example: noticing a pothole. As you walk, the neocortex predicts the feel of each footstep. When its predictions match, you notice nothing. If your left foot comes down and finds no ground, the mismatch makes you look at once.

Key Takeaways

1. The neocortex is one repeated microcircuit, and its regions differ by input, not by algorithm.
2. Filling in, one at a time, and can’t unsee all follow from perception being inference: you perceive a simulation, not raw input.
3. Generative models learn to recognize by generating, with no labels. Modern AI (StyleGAN, GPT-3) shows the idea works.
4. Hallucination, dreaming, and imagination are the generative model running without sensory constraint.
5. The neocortex’s key gift is a world model that can be explored apart from the real world. That is imagination, the third breakthrough.

Connects To

• Ch 4 and Ch 6: earlier forms of prediction (reflexes in bilaterians, reward prediction by the critic).
• Ch 7: supervised learning and auto-associative networks in the vertebrate cortex, which generative models go beyond.
• Ch 12: what early mammals did with the simulation (vicarious trial and error, counterfactual learning, episodic memory).
• Ch 13: the frontal neocortex uses the same microcircuit to model the self, and attention is the same mechanism as the duck-rabbit switch.
• Ch 14: the motor cortex applies the same circuit to the body.