The Stilwell Brain

TL;DR

Emergence is demonstrated by mapping local on/off rules across layered “neurons,” producing a global decision (digit recognition) without any single person holding the full answer.

Briefing Cornell Notes

Briefing

A crowd of hundreds of people can be arranged to behave like a simplified visual brain—processing a drawn digit in real time and using “inhibition” to settle on a single best guess. The experiment turns the idea of emergence into a literal, human-scale neural network: each participant acts like a neuron, firing or staying silent based on whether the “neurons” in the layer ahead of them are active. The result is a working model of how layered visual processing can recognize shapes, even though no individual person knows the digit being drawn.

The core concept starts with emergence: intelligence and emotion are treated as properties that can arise when many simple units connect and coordinate. The transcript then grounds that abstraction with two references. First is the “wisdom of the crowds” Bean Jar experiment, where averaging many wrong guesses lands close to the truth because errors cancel out. Second is the “China Brain” thought experiment—recruiting all people in China as neurons to ask whether a mind could emerge at massive scale. Neither becomes literal here; instead, the project scales the neural-network idea down to a manageable crowd and focuses on one task: digit recognition.

To build the human neural network, Michael heads to Stilwell, Kansas, and recruits Chris Eliasmith, director of the Center for Theoretical Neuroscience at the University of Waterloo, known for SPAUN, a large-scale brain simulation using millions of simulated neurons. The demonstration borrows the structure of early visual cortex processing. A 25-pixel input grid is mapped onto 25 “retina” participants. Higher layers then detect increasingly complex features: V1 identifies simple line patterns, V2 combines those features into angles, V4 builds angle combinations into partial shapes, and the infratemporal cortex (IT) contains ten neurons—one per digit 0–9—that fire only when the corresponding V4 patterns appear.

Logistics matter because each person must have a clear line of sight to the neurons they “connect” to. Participants wear color-coded shirts for each layer and receive bib numbers so the system can be debugged if a specific “neuron” misbehaves. On the football field, Michael draws digits on the grid; retina participants stand and signal “firing” when their assigned pixel contains ink. The network then propagates activity forward through the layers until IT selects the digit.

The system succeeds after an early failure caused by misassigned pixels, then demonstrates robust recognition. An 8 is correctly identified quickly, a 1 is recognized even when extra noise is added, and a 7 is correctly inferred when the drawing is altered with a line and dot. The strongest test comes when Michael fills every cell—forcing the retina to fire everywhere. Instead of collapsing into chaos, inhibition and feature-counting push the network toward a single output: the model still guesses 8, the digit whose structure best matches the strongest competing patterns.

By the end, the crowd has functioned as a simplified living model of visual computation: layered feature detection plus inhibitory competition can turn noisy inputs into a stable decision. The transcript frames the success as proof that collective processing can be observed directly—and hints at how vastly more powerful real brains could be when scaled from hundreds of “neurons” to billions of biological ones.

Cornell Notes

The experiment builds a human-scale neural network that recognizes handwritten digits by mimicking visual cortex layers. A 25-person “retina” receives a 25-pixel grid input and signals on/off firing based on whether ink appears in their pixel. Signals then propagate through V1, V2, V4, and the infratemporal cortex (IT), where IT contains ten digit-specific units that fire only when the right higher-level feature patterns appear. Inhibition helps suppress competing interpretations, so the system can still pick the correct digit even when the input is noisy or when every pixel is filled. The demonstration matters because it makes emergence and feature-based vision tangible in real time.

How does the crowd function as a neural network, even though no individual person knows the digit?

Each participant is assigned to a “layer” and a specific role. The 25 retina participants correspond to the 25 squares of a 5×5 grid; they signal firing (standing and waving a flag) only if their square contains ink. V1 participants watch a small set of retina signals and fire only when their assigned retinal neurons all fire, revealing line features. V2 and V4 similarly require specific combinations of earlier-layer firings (angles and then angle combinations). IT then interprets the resulting pattern by firing only for the digit whose expected V4 pattern appears. Because each person only follows local rules—stand if your inputs are active—knowledge of the whole digit emerges from the network’s connections.

What does “inhibition” do in the digit-recognition model?

Inhibition acts like a competition mechanism that suppresses incorrect or weaker interpretations. The transcript gives a concrete example: if V4 indicates both a 6 and an 8, the 8 can “outrank” the 6 because the 8 has more matching features. In the physical setup, IT neurons use a ball-in-tube mechanism: each IT neuron adds balls when it sees the neurons it watches fire, but if it is inhibited by a competing IT neuron that has more balls, it stops firing. This is why the system can settle on one output even when multiple patterns partially match.

Why did the first attempt fail, and what does that reveal about the system?

The first run stalled after V2: no V4 or IT neurons activated, meaning the signal never reached the higher layers. The likely cause was that pixels were handed out to the wrong retinal neurons—so the early feature detections never formed the correct line/angle patterns required by V2 and V4. The failure highlights that the network’s correctness depends on precise mapping between input pixels and the corresponding “retina” participants, not just on the general layer logic.

How is the digit input represented, and how does that connect to visual processing layers?

Michael draws a numeral onto a 25-pixel grid. Each pixel corresponds to one retina participant. The model then imitates a simplified visual pipeline: V1 detects basic line features; V2 detects combinations of lines that form angles; V4 detects angle combinations that begin to form the number’s shape; IT contains ten neurons (0–9) that fire when the corresponding V4 pattern appears. The demonstration explicitly skips color-sensitive processing (V3) and higher-level steps (V5 and V6) because the task is limited to black-and-white digit recognition.

What happened when every pixel was filled in, and why was the output still a single digit?

When Michael filled every cell, all retina neurons fired. The expectation was that the network might produce a messy or arbitrary result, but inhibition and feature matching still drove the system to a single guess: 8. The transcript frames this as the network effectively “single[ing] all of that mess down into just one guess,” because the digit with the most compatible features (8) dominates the competition among IT units.

What evidence suggests the crowd model can handle noise in the input?

After initial debugging, the system correctly identified a 1 even when Michael added an extra line and a dot—introducing extra markings not part of the intended digit. It also recognized a 7 when Michael modified the drawing with a line through it and additional marks. These tests indicate that the layered feature-detection plus inhibitory competition can tolerate some input distortion rather than requiring perfect, noise-free digits.

Review Questions

What specific conditions must be met for a V1 neuron to fire, and how does that constraint shape what V2 can detect?
How does the ball-and-tube inhibition mechanism determine which IT neuron stops firing, and what does that imply about “winner-take-most” behavior?
Why does misassigning pixels to retina participants prevent higher layers (V4 and IT) from activating?

Key Points

1
Emergence is demonstrated by mapping local on/off rules across layered “neurons,” producing a global decision (digit recognition) without any single person holding the full answer.
2
A 25-pixel input grid is converted into firing signals by 25 retina participants, each responsible for one square’s ink/no-ink status.
3
Layered feature detection is implemented as a cascade: V1 detects lines, V2 detects angles, V4 detects angle combinations, and IT selects among digits 0–9.
4
Inhibition provides a competition mechanism so the network can suppress competing interpretations and choose one stable output.
5
Precise logistics—especially correct pixel-to-retina assignment and clear line-of-sight connections—are critical; early misassignment caused the network to “die” after V2.
6
Even extreme noise (filling every pixel) can yield a correct single guess because feature matching and inhibition favor the digit with the strongest compatible pattern (8).

Highlights

The crowd network recognized digits in real time by treating people as neurons that fire only when their assigned inputs are active.

A first attempt failed because pixels were distributed to the wrong retinal participants, preventing the signal from reaching V4 and IT.

When every pixel was filled, the system still converged on 8—showing inhibition can turn overwhelming input into a single decision.

The model uses a simplified visual pipeline: retina → V1 (lines) → V2 (angles) → V4 (shape fragments) → IT (digit units).

Topics

Emergence
Neural Networks
Visual Processing
Inhibition
Human-Scale Simulation

Mentioned

Michael Stevens
Chris Eliasmith
Lawrence Davis
Ned Block
Jack Treynor
Frank Rosenblatt
Jeff
Brian
IT
V1
V2
V4
V3
V5
V6
SPAUN