Fruit Fly Special 3: A Fruit Fly Brain in Minecraft — A Nervous System Driven by Its Wiring Diagram Alone

What does this game do? — Three ingredients glued together

Put a fruit fly mob into Minecraft, and run an entire real fruit fly nervous system inside its head. What the mob sees, smells, and tastes stimulates real sensory neurons, and the activity of real descending neurons and motor neurons is read out to move the mob. Each fly runs its own brain on its own thread in real time.

Ingredient 1 · Wiring diagramThe connectome of an entire male fruit fly nervous system

The whole central nervous system: the brain plus the ventral nerve cord (the insect's spinal cord). It uses the public data (neuPrint male-cns:v1.0) from the paper by Berg et al. published in Cell on September 3, 2026. The mod was built on the very day the paper came out.

Ingredient 2 · Neuron modelEvery neuron is the same "leaky integrate-and-fire" part

It follows exactly the method Shiu et al. (2024) Nature used to simulate the whole fruit fly brain. No values are tuned per neuron; the synapse count of a connection is its connection strength.

Ingredient 3 · Body and worldThe Minecraft world and the mob's body

Apples, cakes, spider eyes, rain, and charging players become sensory input, and the brain's commands become the mob's movements such as walking, eating, jumping, and grooming. This part was built by hand.

Minecraft screen: a flying fruit fly mob, the nervous system map on the left, the Neuroscope panel on the right — **Actual game screen.** In the center is the fruit fly mob (Fly-1 ♂); on the left is the B-key "brain map" (top: optic lobes and brain, bottom: ventral nerve cord); on the right is the H-key "Neuroscope" (motor channels, firing rates of neuron groups, spikes per tick, retina). The behavior observation section below is modeled on this panel.

A structure building the connectome out of stained-glass blocks: the brain on the left, the ventral nerve cord on the right — **A connectome built from blocks.** The `/flybrain build` command places the 141,781 neurons whose cell-body positions have been reconstructed as stained-glass blocks colored by region. When a neuron in a linked mob fires, its block flashes as a sea lantern.

In one line

If the previous two pages saw "adjusting connection strengths through experience" as the heart of neural networks, this mod tests "how far you get when the connection strengths are taken straight from the wiring diagram." There is no learning. Few numbers are set by people. The most important is the global gain of 0.65 multiplied into every connection; besides that there is a 0.25× factor on Kenyon cell input and the rule "drop connections with fewer than 5 synapses."

The fruit fly research behind this game

We grouped the studies cited in the repository's README and reference document (REFERENCE.md) by role. Two papers are central: what is connected comes from Berg et al. (2026), and how it is run comes from Shiu et al. (2024).

Core ① · Wiring diagram

Berg et al. (2026) Connectome of the entire male fruit fly central nervous system

The brain and ventral nerve cord were imaged as a single electron-microscopy volume, and every neuron and synapse was traced. It is also the first study to compare male and female brains at the level of individual synapses.

Cell 189(18):5504–5526 · doi:10.1016/j.cell.2026.08.015

Core ② · Neuron model

Shiu et al. (2024) Computational model of the whole fruit fly brain

They ran every neuron in the FlyWire female brain wiring diagram as a leaky integrate-and-fire part, predicted sensorimotor pathways such as sugar taste → proboscis extension, and confirmed them experimentally. The mod's neuron equations and numbers are taken straight from this paper.

Nature 634:210–219 · doi:10.1038/s41586-024-07763-9

Data tool

Plaza et al. (2022) neuPrint

A public tool for querying connectomes. All the neuron names in the mod (DNp01, MN9 …) are neuPrint cell-type names, so you can look up any neuron on screen directly on the web.

Front. Neuroinform. 16:896292

Companion papers on the sensory side

Visual system · taste connectome

Nern et al. (2025) full inventory of the visual system, Hoeller et al. (2026) visual pathways, Tastekin et al. (2026) taste-feeding connectome. Used to decide which sensory neurons receive the game input.

Nature 641 · 2 papers in Cell 189(18)

Evidence on the motor side

Descending neurons and behavior

Walking and stopping (Sapkal 2024), turning (DNa02), the giant fiber and looming escape (von Reyn 2014, Ache 2019), landing (Ache 2019), antennal grooming (Hampel 2015), wingbeat strength (Namiki 2022). These are the basis for the rules that read neuron activity as the mob's movement.

Nature · Nat Neurosci · eLife · Curr Biol

Ventral nerve cord · sex-specific circuits

MANC connectome and sexually dimorphic circuits

Male ventral nerve cord studies by Takemura et al., Marin et al. (2024), and Cheong et al. (2026), and the study of male-specific neurons (pC1, P1) by Rubin et al. (2026). They underlie the courtship-song connections, but this feature has not been demonstrated yet.

eLife 13 · Curr Biol (2026)

Why the numbers differ slightly. The paper's abstract counts 166,700 proofread neurons. The mod simulates all 176,422 entries loaded from the neuPrint dataset, and builds the 141,781 of them whose cell-body positions have been reconstructed into the block structure. The difference seems to come from different counting criteria.

Relation to the earlier page. The FlyWire (2024) connectome cited on page ① is the very female brain wiring diagram that Shiu et al. (2024) used. This mod applies the same method to a larger male nervous system (brain + ventral nerve cord).

From the brain to the ventral nerve cord — the path signals take in the game

To the 3D brain from the previous page we added the ventral nerve cord (the long lump at the bottom), the neck, and the proboscis, because this mod runs not just the brain but also the motor neurons that move the legs and wings. Pick a scene and a green signal follows the actual connection pathway. This is a schematic model, not real anatomical data.

Scene

Region info

What one neuron computes — pouring water into a leaky bucket

All 170,000 neurons are this one part. When input spikes arrive the voltage rises, and if left alone it leaks back down to the resting voltage. Once it crosses the threshold (−45 mV) it fires a spike and immediately resets to −52 mV. It is page ①'s "super-simple model" y = [w·x ≥ θ] with time and leak added.

Input neuron firing rate

Synapses per connection

Global gain

The artificial neuron on page ①

Sum input × weight; output 1 if it crosses the threshold, otherwise 0. There is no time, so the answer appears the instant the input arrives. Weights are set by learning.

This mod's leaky integrate-and-fire neuron

Input builds up over time and leaks away. So "how often, and how many at once" inputs arrive matters. The weight is fixed at the wiring diagram's synapse count × 0.275 mV × gain 0.65.

How are excitation and inhibition decided? — The sign also comes from the wiring diagram

The sign of a connection is set by the neurotransmitter the neuron releases. Acetylcholine, dopamine, serotonin and others are excitatory (+); GABA, glutamate, and histamine are inhibitory (−). All outputs of one neuron have the same sign (Dale's principle). The result: 116,390 excitatory and 60,032 inhibitory neurons. The "inhibitory synapses are negative weights" line in page ①'s summary table is set here directly by the wiring diagram.

Behavior observation dashboard — give a stimulus, watch the neurons, read the behavior

Modeled on the mod's "Neuroscope" (H key) screen. Pick a scene: sensory neurons are stimulated, the signal passes through intermediate neurons, and the firing rates of descending and motor neurons go through the decoder's priority ladder to become behavior. The pale band on each bar is the actual measured range from the repository's validation record (VALIDATION.md).

Scene

Minecraft world→ Sensory neurons · Poisson spikes→ Brain · 50 ms of computation→ Descending/motor neurons fire→ Decoder · priority→ Mob behavior

Firing rate of neuron groups (Hz)

Decoder priority ladder (leftmost is strongest)

How does the decoder pick a behavior?

It builds 21 command channels (forward, turn, backward, stop, jump, eat, groom …) by summing the firing rates of neuron groups. When two or more are on at once, they win in this order: escape > landing > flight > sudden stop > stop > backward > forward > eat > groom > song. Once entered, a behavior stays for at least 250 ms and exits only when its signal drops below half the threshold. Which neurons turn on is decided by the wiring diagram, but this ladder and these thresholds were set by people.

The most important control knob — tuning the global gain

Shiu et al.'s 0.275 mV per synapse was fitted to the female brain (FlyWire). The male nervous system has more synapses per neuron, so using that value as is makes the whole network overexcited. So a single gain multiplied into every connection was re-tuned using the same method as the paper.

Gain

In this mod

They picked the smallest gain that satisfies all three conditions: ① sugar input reliably turns on the proboscis motor neuron; ② the grooming and escape pathways are stable; ③ the Kenyon cells of the learning center stay quiet. At 0.55 eating is weak; at 0.75 activity runs away.

In deep-learning terms

It's like leaving individual weights alone and choosing a hyperparameter, such as the initialization scale, through validation experiments. The 0.25× factor for Kenyon cells corresponds to a correction applied to one specific layer. And if it's too large, activations blow up as they pass through layer after layer — a "runaway" just like in deep neural networks.

What comes from the wiring diagram, and what was made by people?

The best thing about this repository is that it draws the line itself. The README states that "the point of this project is being honest in the last column (source)." Each behavior is tagged emergent (arising on its own from the wiring diagram), hand-built (a stand-in made by people), or mixed.

Mind the scope of "emergent." Here, emergent means "it came out of this computer model without hand-written rules." It is not a result confirmed in real male fruit flies, but the result of automated tests (benches) that run the model alone without the game. Some of the same pathways (sugar → MN9, grooming) were predicted by Shiu et al. (2024) with a female brain model and confirmed in animal experiments.

Behavior	In-game trigger	Pathway in the wiring diagram	Source

What arises on its own from the wiring diagram

Which descending and motor neurons turn on for which sensory pattern · bitterness switching off eating just before the motor stage · left/right asymmetry in turning · the opposition between walking forward and backward · the giant fiber's single-spike jump decision · the choice of which body part to groom

Stand-ins built by people

Retina rendering and brightness adaptation · the odor map and distance falloff · the scale factor that converts firing rate into blocks/second · thresholds and the priority ladder · flight and landing physics · injecting the looming signal directly into LC4 and LPLC2 · the 0.25× factor on Kenyon cell input · the reflex layer that walks toward odors

Known limitations (from the repository docs). Every neuron uses the same values · smelling an odor saturates the antennal lobe, so there is no direction signal · the medulla's motion pathway (T4, T5) is silent · no neuromodulators, neuropeptides, electrical synapses, or spontaneous activity · connections with fewer than 5 synapses were dropped · females also use the male brain · don't trust the absolute firing rates; look only at the pattern of which groups respond.

Try it elsewhere — this mod's design seen through other AI

This mod is a neuroscience project, but its design has a lot in common with today's AI systems. We've mapped out the parallels for people learning or teaching generative AI, and for those bringing AI into the workplace.

In generative-AI terms

Bringing it into your own system — if your work mixes rules and learning

Limits of the analogy. The weights of a large language model aren't designed by people but learned from data, whereas this mod's weights are measurements of a real animal's wiring. They only look alike in that both "just run inference with fixed weights"; where those weights come from is exactly the opposite. And unlike an agent's tool selection, the priority ladder doesn't understand language — it only compares the size of firing rates.

Comparing three projects — same fruit fly, different questions

Page ① is an introduction to neural networks, page ② illustrates the memory retrieval paper, and then there's this Minecraft mod. All three follow "senses → brain → behavior," but they differ in what they keep real and what they simplify.

Where the three projects sit

Pick a shared topic

How to read the map on the left

The horizontal axis is how faithfully the real wiring diagram is used, and the vertical axis is how central it is to modify connections through experience. The previous two pages sit in the upper left, and the mod in the lower right. The upper right — models that learn on top of a real wiring diagram — is a field that has only just begun. A prime example is Lappalainen et al. (2024) Nature, which kept the real connections among 64 cell types of the fruit fly visual motion pathway fixed and used deep learning to train only the unknown neuron and synapse values. That model's predictions matched measurements from 26 studies. It is the very pathway (T4, T5) that is silent in this mod. However, no model yet runs the entire brain and body together while also including learning and neuromodulation.

Side-by-side comparison table

Similarities

Differences

Comparing model architectures — this page's model vs. MLP · RNN · GRU · LSTM · CNN

Press the buttons to switch between diagrams of the six architectures. Green lines are learned connections, dotted lines are fixed connections, and orange arrows are memory carried across time. All diagrams are schematics that illustrate the principle.

All six architectures at a glance

Learned connections / total connections (input 16×16, 3 classes)

Gray bar = total connections, colored bar = learned connections. Deep learning models learn everything; in the fruit fly models most connections are fixed. The trainer below and Code 2 in common section A count the same numbers.

Level of evidence in the last row: Real circuit/measured the model copies a real circuit · Similar structure the connection pattern looks alike, which is not evidence that it performs the same computation · Loose analogy · Pure analogy just a figure of speech to aid explanation, with no biological basis.

The key difference from this page's point of view

Surprisingly, the architecture in the table most like this model is the RNN. Both run a loop over time and have feedback connections. The decisive difference is where the weights come from. An RNN starts random and learns by backpropagation; this model uses wiring measured by electron microscopy as is. In between lies the method of Lappalainen et al. (2024): keep the wiring as measured and train only the unknown values with deep learning.

Fruit fly trainer — teach it with your own drawings and try it right away

Like Teachable Machine: ① collect examples for each class, ② train, and ③ check with new input. All computation happens inside this browser; your drawings and webcam video never leave it. To run the same model in code, use Code 2 in common section A.

① Collect data

② Train

Model

Epochs30

Learning rate0.05

③ Preview

Input

Export

Continue with Code 2 →

Load · Save

Collected samples are saved automatically in this browser and survive a page refresh. Exported data and model JSON files can be loaded back in.

Getting started: press "Fill sample data," or draw on the pad and press "Add drawing" on a class. Click a name to rename it.

Same as Teachable Machine

The flow of create classes → collect samples → train → live preview → export, webcam input, and doing all computation in the browser.

Differences

Teachable Machine uses transfer learning: it trains only a small classifier on top of features pre-extracted by a large image model (MobileNet). This trainer has this page's models learn 16×16 drawings from scratch. So it suits simple shapes and struggles with complex input like photos. In exchange, it shows you the validation accuracy and the number of learned connections directly.

Summary on one page

What the previous two pages showed

Artificial neural networks that borrow the brain's principles solve problems by adjusting connection strengths through experience. The fruit fly's mushroom body likewise changes connections with dopamine to pair odors with outcomes, and that memory becomes behavior by passing through the instinct circuit (lateral horn).

What this mod showed

If you take connection strengths straight from the wiring diagram, eating, rejecting bitterness, jumping, and grooming come out of the right neurons even without learning. On the other hand, parts that need learning and modulation, such as following odors, are still handled by human-made stand-ins.

Key point 1. Structure (wiring) alone produces many behaviors. Much of a fruit fly's innate behavior is engraved in the connections themselves.

Key point 2. Learning and memory (mushroom body, dopamine) are the blank this mod leaves empty. The Kenyon cells are deliberately kept quiet, and there are no neuromodulators. The memory retrieval circuit on page ② is exactly that blank.

Key point 3. All three projects mark "how much of this is real" — the pages with yellow stamps, the mod with emergent/hand-built tags. That's the most important habit when making science easy to see.

Revisit: ① Exploring the Fruit Fly Brain — from senses to recognition · ② Learned memories are retrieved through the instinct circuit

Colab code — run this page yourself

Both codes use only libraries preinstalled in Colab. Click "Download notebook (.ipynb)" and open it in Colab via File → Upload notebook, or copy and paste it into a new notebook cell.

Code 1 · A leaky integrate-and-fire network fixed by the wiring diagram, in NumPyLinked to the "One neuron" and "Behavior" sections · numpy · under 1 s on CPU

It builds fixed weights as synapse count × 0.275 mV × gain 0.65 (sign set by the neurotransmitter), stimulates 20 sensory neurons at 120 Hz, and computes 100 steps of 0.5 ms each. Not a single weight is learned. To use your own data, load syn (the synapse-count matrix) from a real connectome file.

Output (run on 2026-10-04, numpy 2.4 · CPU). The random seed is fixed, so you get the same result whenever you run it.

Code 2 · Five standard deep-learning models vs. three fruit fly models in PyTorchLinked to "Hands-on: architecture comparison and trainer" · torch, numpy, matplotlib · about 1 min on CPU

Trains MLP, CNN, RNN, GRU, LSTM, and three fruit fly models on the same drawing data. The fixed-wiring spiking model (ConnectomeLIF) fixes the connections of its 160 hidden neurons and trains only the readout layer. Cell 8 plots the spike raster.

Output (run on 2026-10-04, PyTorch 2 · numpy 2.4 · CPU). Trained separately with two seeds. Numbers may vary with seed, version, and hardware.

How is it different from ordinary deep-learning code?

Item	Code 1 (fixed-wiring spiking)	Ordinary deep-learning code
Source of weights	Synapse count measured by electron microscopy × 0.275 mV × gain	Start random, then learn
Learning	None (people calibrate only a few values, such as the gain of 0.65)	All connections learned with surrogate gradients
Derivative of a spike	Not needed (there's no learning)	It's a step function, so the gradient is 0 → work around it with a surrogate gradient
Handling time	0.5 ms × 100 steps = 1 game tick	20-step time loop + backpropagation through time
Scale	176,000 neurons, 6.29 million connections (500 in the example)	Typically hundreds to thousands of neurons
Purpose	Predict and explain what a real brain would do	Task performance (accuracy)
Reading the output	A human-made decoder (priority ladder)	The output layer directly gives class scores

View ordinary deep-learning code for comparison (snn_surrogate_gradient.py, PyTorch)

AI prompts — use this code as a starting point

Prompts to paste into an AI assistant such as ChatGPT, Qwen, or Claude. Pressing "Copy" appends the selected code after the prompt. Replace the parts in 〔〕 with your own situation.

Code to attach

An AI's answer is only a starting point. Always verify any suggested papers, figures, or library functions against the original sources and official documentation. If you used AI help for a competition entry or paper, disclose that you did and to what extent, following the rules of the organizers or journal.

Glossary

This is the same dictionary as the yellow dotted underlines in the text. Hover over or tap an underline to see the same explanation.

Analogies — and where they break down

Analogies are a first step to understanding, but they can also plant misconceptions. Fruit fly/brain analogies are marked by strength: Direct there is evidence it actually does the same thing · Partial only some properties are alike · None no known counterpart

Concept	Everyday analogy	Fruit fly/brain analogy	Where the analogy breaks down
Leaky integrate-and-fire	Pouring water into a leaky bucket: when it overflows, empty it and fill it again	Direct A real neuron's membrane voltage builds up with input, leaks away, and fires a spike at threshold	Real neurons vary in shape and ion channels. This model uses the same values for every neuron
Wiring diagram = pretrained weights	Getting a map someone else made and only using it to find your way	Partial Synapse counts measured by electron microscopy are used directly as connection strengths	Generative AI weights are learned from data; the wiring diagram is measured. Their origins are complete opposites
Gain tuning	Turning a single master volume knob on a stereo	None	A real brain has no global gain knob. It's a value people chose to fit the male nervous system
Reflex layer (stand-in)	A human grabbing the wheel when a self-driving car hesitates	None	The reflex layer isn't a biological circuit but rules written by people. The [REFLEX] tag on screen marks those moments
Emergence	Each ant is simple, yet the whole colony behaves in complex ways	Partial No rules were added, yet feeding and escape come out of the wiring diagram	"Emergence" is a result inside this model. It hasn't been confirmed in real male fruit flies

About the Author and References

E-1 About the author · E-2 Academic references · E-3 Image credits

E-1 About the Author

Sangsun Ahn (CEO of M-Robo Inc.)

After working at Korea Electrical Safety Corporation and Maeil Business Newspaper, he founded M-Robo Inc., a spin-off venture of Maeil Business Newspaper, where he currently serves as CEO. M-Robo started as an in-house venture of a news company and grew into a media venture: it provided a robot journalist (StockBot) to Kukmin Ilbo and carried out many projects for the Korea Internet & Security Agency, the Korea Press Foundation, the Korea Communications Agency, and others. It now runs its own AI news service, Money-Robo. While serving as CEO, he won the Asia Open Data Hackathon and twice received the Minister of the Interior and Safety Award for contributions to data entrepreneurship, and he is the inventor and registrant of five patents in artificial intelligence and blockchain. He is the author of “AI Crime and Self-Defense” (in Korean), and has developed e-learning content on fintech, statistics, and artificial intelligence for the Online Lifelong Education Institute of KOREATECH (Korea University of Technology and Education). He holds a PhD in Economics (Dongguk University, microeconomics and mathematical economics). He teaches data and AI courses at several universities, including Kookmin University and Hoseo University. Full profile

Courses

Sangsun Ahn (2025). “Data Science for Business” lecture notes (in Korean). Department of AI Convergence, Open Cyber University of Korea.
Sangsun Ahn (2025). [K-Digital] Natural Language Processing with AI (in Korean). STEP (Smart Training Education Platform), KOREATECH Online Lifelong Education Institute.
Sangsun Ahn (2024). [K-Digital] Fundamentals of Fintech Data Analysis (in Korean). STEP (Smart Training Education Platform), KOREATECH Online Lifelong Education Institute.

STEP course links: link.step.or.kr/1qyVoJ · link.step.or.kr/M9na9N · link.step.or.kr/brqEyc
If the links don't work because of a site redesign, please go to www.step.or.kr.

Papers and Presentations

Sangsun Ahn (2025). A study on household expenditure prediction models using the Korean Retirement and Income Study panel data: An AI model-based approach (in Korean). “Future Society”, 16(2). Full text
Sangsun Ahn (2025). A study on household expenditure prediction models using the Korean Retirement and Income Study panel data: An AI approach (in Korean). Presentation at the KReIS Panel Conference (2025-05-13), National Pension Research Institute, National Pension Service. Slides
Sangsun Ahn (2025). A study on algorithms for detecting suspected drug transactions from SNS text: Focusing on pattern and context analysis (in Korean). “Future Society”, 16(1). Full text
Sangsun Ahn (2024). Statistics-based countermeasures against adversarial data for securing anomaly detection models in finance: Empirical analysis and simulation results (in Korean). “Future Society”, 15(3). Full text
Sangsun Ahn (2024). An AI-based solution for detecting SNS posts based on virtual character traits (in Korean). Poster at the 2024 Summer Newmedia Party (2024-06-19), Seoul Media Institute of Technology.
Sangsun Ahn (2024). Machine learning models using industrial accident panel data (in Korean). Poster at the 10th Panel Study of Workers' Compensation Insurance Conference (2024-05-17), Labor Welfare Research Institute, Korea Workers' Compensation & Welfare Service.
Sangsun Ahn (2024). Analysis and forecasting of virtual asset time-series data: Comparing time-series and AI models, with suggested improvements (in Korean). “Future Society”, 15(1). Full text

E-2 Academic References

[R1] Berg, S., Beckett, I. R., Costa, M., … (2026). Sexual dimorphism in the complete Drosophila male central nervous system connectome. Cell, 189(18), 5504–5526. doi:10.1016/j.cell.2026.08.015 — hero, "Related research"
[R2] Shiu, P. K., Sterne, G. R., Spiller, N., … Scott, K. (2024). A Drosophila computational brain model reveals sensorimotor processing. Nature, 634, 210–219. doi:10.1038/s41586-024-07763-9 — "One neuron", "Tuning the gain"
[R3] Lappalainen, J. K., Tschopp, F. D., Prakhya, S., … Turaga, S. C. (2024). Connectome-constrained networks predict neural activity across the fly visual system. Nature, 634, 1132–1140. doi:10.1038/s41586-024-07939-3 — "Comparing three projects"
[R4] Plaza, S. M., Clements, J., Dolafi, T., … (2022). neuPrint: An open access tool for EM connectomics. Frontiers in Neuroinformatics, 16, 896292. doi:10.3389/fninf.2022.896292 — "Related research"
[R5] blendi-remade/fly-brain-minecraft (2026). GitHub repository: README, docs/REFERENCE.md, docs/VALIDATION.md, PROVENANCE.md. Code MIT, connectome-derived data CC BY 4.0. github.com/blendi-remade/fly-brain-minecraft — this whole episode (the mod's numbers and validation records)

Sources for the numbers and facts in footnotes [1]– of the main text

No.	Content	Value	Source
[1]	Number of simulated neurons, connections, and synapses	176,422 / 6,287,749 / 90.3 M	fly-brain-minecraft README, PROVENANCE.md, VALIDATION.md §0 (neuPrint `male-cns:v1.0`, only connections with 5 or more synapses)
[2]	Number of retinal columns, 50 ms tick, no learning, neuron equation values	877 + 892 columns · τ_m 20 ms · τ_s 5 ms · −52/−45 mV · refractory period 2.2 ms · 0.275 mV × 0.65	docs/REFERENCE.md "How it works", Shiu et al. (2024)
[3]	Firing-rate ranges, active neurons, and real-time factor in the behavior dashboard	Ranges in the table	docs/VALIDATION.md §1–8 (automated tests run without the game, 2026-09-03)
[4]	Comparing gain 0.55 / 0.65 / 0.75	§9 table	docs/VALIDATION.md §9
[5]	Number of neurons in the connectome paper	166,700 (proofread and annotated)	Berg et al. (2026) Cell 189(18):5504–5526
[6]	Whole-brain leaky integrate-and-fire model	—	Shiu et al. (2024) Nature 634:210–219
[7]	Visual system model with values learned by deep learning on top of the real wiring diagram	64 cell types, matches 26 experimental studies	Lappalainen et al. (2024) Nature 634:1132–1140 · doi:10.1038/s41586-024-07939-3
[8]	Two game screenshots	—	fly-brain-minecraft docs/media (MIT License)

E-3 Image Credits

Two game screenshots: fly-brain-minecraft repository docs/media (MIT License, © 2026 Blendi Remade / fal.ai). Minecraft is a trademark of Mojang/Microsoft
Character art in the closing scene: Deep Learning A to Z lab characters

TIPs for Educators

A guide for instructors, on-the-job training (OJT) leads, and tutors. Times are estimates; adjust them to your actual class.

Audience	Learners who have finished Special 1 and 2. Practitioners interested in brain simulation and digital twins
Prerequisites	Special 1's ② Perceptron; Special 2's fixed pathway + learned correction
Learning goals	① Explain the leaky integrate-and-fire neuron equation in words ② Tell apart behavior that emerges on its own from the wiring diagram and stand-ins built by people ③ Compare the strengths and limits of a model with almost no learned connections against deep learning
Estimated time	Reading 50–60 min · animation demos 25 min · Colab practice 30 min (Code 1: 10 min; Code 2: 2 min to run + 15 min to interpret)

Three lesson plans

Lecture, 50 min

Concept-focused

Intro 5 min (game screens) → One neuron 15 min → Watching behavior 15 min → Real vs. stand-in 10 min → Check questions 5 min.

Workplace OJT, 90 min

System-design-focused

Watching behavior 15 min → Rewrite Real vs. stand-in as a "model vs. rules" table for your own system 25 min → Try transferring it 15 min → Run and interpret Code 2 25 min → Prompt cards 10 min.

1:1 tutoring

Pace by level

Beginners: just What is it, One neuron, and Watching behavior. Intermediate: derive the "162 simultaneous synapses" from the equation in One neuron, and change the gain in Code 1 to compare spike counts.

Teaching points by section

Section	Question to pause and ask	Hands-on tip
What is it	"Can a neural network that doesn't learn be useful?"	Start with the two game screenshots
One neuron	"Why do spike times differ even with the same input?"	Change the synapse count and gain and watch the output firing rate
Watching behavior	"At which stage does feeding stop when you add bitterness?"	Press Sugar, then Sugar + bitter, in that order
Tuning the gain	"Why does it run away at 0.75?"	Compare the grooming graphs at 0.65 and 0.75
Real vs. stand-in	"When [REFLEX] appears, who's driving?"	Replay the apple-odor scene
Comparing three projects	"What sits in the upper right (real wiring diagram + learning)?"	Point to the Lappalainen (2024) dot on the map
Practice: Trainer	"How many % can 483 learned connections get right?"	Alternate training the fixed-wiring spiking model and the MLP

Top 5 Common Misconceptions

Misconception	Correct explanation	Scene to show
A connectome alone fully reproduces the brain	Neuromodulators, electrical synapses, and neuron-specific properties are missing. The odor pathway saturates, so a stand-in takes over	The limitations list in Real vs. stand-in
The fly in the game learns	There are 0 learned weights. Only a few values, such as gain 0.65, were set by people	Hero numbers, Tuning the gain
The firing-rate numbers are those of real fruit flies	They're benchmark results from the model; don't trust the absolute values, look only at the patterns	The faint bands in Watching behavior
The emergent behaviors were confirmed in experiments on real fruit flies	They're results inside this model. Only some pathways were confirmed in animal experiments by other studies	The paragraph on the scope of "emergence" in Real vs. stand-in
Spiking neural networks can't learn	In machine learning they're trained with surrogate gradients. That this mod doesn't learn is a design choice	Structure comparison, Code 2

Check questions (click to reveal explanations)

Q1. For a connection of 30 synapses, how many mV of current g does one input spike add? (0.275 mV × gain 0.65)

30 × 0.275 × 0.65 ≈ 5.36 mV.

Q2. With a resting voltage of −52 mV and a threshold of −45 mV, if one spike raises the membrane voltage by about 0.1575 × g, how large must g be to make the neuron fire on its own?

7 mV to threshold. 7 / 0.1575 ≈ 44.4 mV. At gain 1.0 that takes about 162 synapses at once; at gain 0.65, about 249.

Q3. In the sugar + bitter scene, the proboscis motor neuron (MN9) is at 0 Hz, yet the sweet second-order neuron is on. At which stage does the inhibition happen?

At the stage just before the motor output. The bitter brake (Scapula) blocks the feeding-command pathway.

Q4. What happens in the grooming scene if you raise the gain to 0.75?

Within 200 ms, spikes per tick grow fivefold, Kenyon cells wake up at 39 Hz, and the resulting inhibition even shuts off the grooming command neurons (runaway).

Q5. In Code 2, the fixed-wiring spiking model scored 0.75–0.81 on the test set with 644 learned connections. How would you interpret this compared with the MLP (16,708 connections, 0.88–0.91)?

Learning only about 4% of the connections still gets a fair amount right, but because it can't learn the representation itself, it scores lower than the MLP. Also mention that with so little data (32 test samples), the difference swings by ±0.1.

Colab assignments

Basic

In Code 1, change the gain from 0.65 to 0.55 and 0.75, report the number of neurons that fired and the average firing rate in a table, and interpret it in two sentences, linking it to the runaway in the "Tuning the gain" section.

Challenge

In Code 2's ConnectomeLIF, change every inhibitory neuron sign (−0.5) to +0.18 (removing inhibition) and retrain. Compare the spike raster (cell 8) and accuracy across two seeds, and write about what inhibition does in a neural network.

Adjusting Difficulty and Learning More

Beginners	For the "One neuron" equation, reading just the "In plain words" part is enough. Make sure Watching behavior and Real vs. stand-in are solid
Advanced	Have students research and present methods that learn only some values on top of a wiring diagram, like Lappalainen et al. (2024). Check the papers they research in the original
Learn more	Sangsun Ahn, “Data Science for Business” lecture notes (data science overview, in Korean) · STEP [K-Digital] Natural Language Processing with AI · STEP [K-Digital] Fundamentals of Fintech Data Analysis
Instructor materials	Grading criteria for the check questions and sample answers for the assignments are on the Answer Key for Educators page (Special 3)

What does this game do? — Three ingredients glued together

In one line

The fruit fly research behind this game

Berg et al. (2026) Connectome of the entire male fruit fly central nervous system

Shiu et al. (2024) Computational model of the whole fruit fly brain

Plaza et al. (2022) neuPrint

Visual system · taste connectome

Descending neurons and behavior

MANC connectome and sexually dimorphic circuits

From the brain to the ventral nerve cord — the path signals take in the game

Region info

What one neuron computes — pouring water into a leaky bucket

How are excitation and inhibition decided? — The sign also comes from the wiring diagram

Behavior observation dashboard — give a stimulus, watch the neurons, read the behavior

Firing rate of neuron groups (Hz)

Decoder priority ladder (leftmost is strongest)

How does the decoder pick a behavior?

The most important control knob — tuning the global gain

What comes from the wiring diagram, and what was made by people?

Try it elsewhere — this mod's design seen through other AI

In generative-AI terms

Bringing it into your own system — if your work mixes rules and learning

Comparing three projects — same fruit fly, different questions

Where the three projects sit

Pick a shared topic

How to read the map on the left

Side-by-side comparison table

Similarities

Differences

Comparing model architectures — this page's model vs. MLP · RNN · GRU · LSTM · CNN

All six architectures at a glance

Learned connections / total connections (input 16×16, 3 classes)

The key difference from this page's point of view

Fruit fly trainer — teach it with your own drawings and try it right away

Summary on one page

Colab code — run this page yourself

How is it different from ordinary deep-learning code?

AI prompts — use this code as a starting point

Glossary

Analogies — and where they break down

About the Author and References

E-1 About the Author

Courses

Papers and Presentations

E-2 Academic References

Sources for the numbers and facts in footnotes [1]– of the main text

E-3 Image Credits

TIPs for Educators

Three lesson plans

Concept-focused

System-design-focused

Pace by level

Teaching points by section

Top 5 Common Misconceptions

Check questions (click to reveal explanations)

Colab assignments

Basic

Challenge

Adjusting Difficulty and Learning More

Great work!