What you will learn
Follow the project ladder alongside prerequisite parts; each project needs evidence, not a completion click.
Before you begin: Foundations: matter, mathematics & computation
START WITH A QUESTION
Picture this.
You do not need a million neurons to build a meaningful project. You need a question, a working reference and a result you can explain. A small system with clear tests is a better launch point than a large system whose behavior is mysterious.
- Reference implementation
- A simple version used to check the behavior of later implementations.
- Regression test
- A repeatable check that a previously verified behavior still holds.
- Payload
- The data itself, excluding additional container and runtime overhead unless stated.
Read for the idea first. Work through the numbers next. Try answering the questions before opening the explanations.
Build a project that grows one question at a time
Begin with the passive-neuron reference from Part 6. Keep the analytic comparison as a test. Add threshold and reset with a declared event-time convention. Connect three units using the orientation check from Part 4. Feed a synthetic visual cue, then decode the output into a simulated scalar command. At each step, save a small artifact: a table, trace, graph or log. You are building a chain of explanations, not merely accumulating components.
Use a modest capstone
A feasible capstone is a simulated cue-to-stop pipeline. Generate a sequence containing quiet periods, true cues and distracting cues. Compare a direct threshold baseline with a stateful detector. Decode the result into a bounded request for the ideal plant from Part 13. Measure false alarms, missed cues and response delay before adding another model. The synthetic setup isolates learning goals; it does not reproduce a fly brain or establish a physical safety system.
Keep the experiment rerunnable
Record the input sequence or its generator seed, all parameters, units and update order. Separate code from result files and keep an expected small reference output. A seed helps reproduce a random sequence in a specified environment; it does not make one run representative of every case. Test a quiet input, a pulse, a boundary-threshold value and a wrong-orientation case. If an optimization changes results, investigate whether it changed the numerical method or event ordering.
Estimate resources, then measure
For N nodes with mean outgoing degree d, the edge count is N*d. Multiply by a stated bytes-per-edge assumption to estimate payload storage. A dense float32 weight matrix uses 4*N² bytes. These estimates omit allocator and framework overhead unless explicitly included. Also budget neuron states, delay queues, saved activations and gradients if training. Sparse storage and fast execution are different claims. Profile the small correct implementation before replacing it with a harder-to-inspect one.
SEE THE RELATIONSHIPS
From structure to behavior.
MAKE THE IDEA PRECISE
The mathematics.
| Symbol | Meaning |
|---|---|
| N | Number of neurons |
| d | Chosen average outgoing degree |
| M | Number of directed edges |
| b_edge | Illustrative bytes per stored edge |
| b_weight | Bytes per dense weight |
Build the equation, step by step
- For N nodes with average outgoing degree d, the number of directed edge records is N d when d is calculated over the same graph.
- Multiply M by bytes per edge record for a sparse payload estimate. Multiply N squared by bytes per weight for the dense payload estimate.
- These estimates omit allocator, index, metadata and runtime-state overhead unless included explicitly. Sparse storage reduces payload when M is small, but runtime speed also depends on access patterns and hardware.
WORK THROUGH THE NUMBERS
Worked example
For 1000 neurons with average degree 10, there are 10,000 edges. At an illustrative 8 bytes per edge, the edge payload is about 78.1 KiB. A float32 dense matrix uses about 3.8 MiB. Neither number includes all runtime overhead.
A small Python experiment.
Python 3 · Standard library only · Synthetic educational example. Run this in your own notebook or a local file.
N, degree = 1000, 10
sparse = N*degree*8
dense = N*N*4
print(f"edge payload = {sparse/1024:.1f} KiB; dense = {dense/1024**2:.1f} MiB")Read the code as a story
The sparse estimate assumes an 8-byte edge payload; it is not the memory occupied by a Python tuple. Dividing by 1024 gives KiB and by 1024² gives MiB. The dense estimate assumes four bytes per weight. This code estimates storage only; it performs no neural simulation or timing benchmark.
Expected output
edge payload = 78.1 KiB; dense = 3.8 MiB
Change one quantity, predict the result, and then run the code again. Explain any difference before adding complexity.
Test your understanding.
Try it: Include one float32 membrane value per neuron and estimate the extra storage. What other arrays would a trained SNN require?
Self-check: Does a smaller sparse representation always execute faster?
No. Irregular indexing and scattered memory access can outweigh fewer arithmetic operations, while dense kernels can exploit specialized hardware. Measure representative workloads and include conversion costs; a storage estimate is not a speed benchmark.
Where this model stops: The example estimates storage, not throughput or end-to-end simulator memory.
A research question: Which optimization changes runtime while preserving the numerical behavior you care about?
Record evidence in your mastery worksheet →PREDICT / TRY / EXPLAIN
Write the capstone's first acceptance test
Use paper, a calculator or the Python example. Write your prediction before checking the result.
- Choose one deterministic input sequence and specify the expected detector behavior before coding.
- Add a quiet sequence and a distracting sequence. Decide what constitutes a false alarm and a missed cue.
- For N = 1000, add one float32 membrane value per neuron to the edge-payload estimate. List at least two additional allocations needed by a richer simulator.
Compare with a worked solution
The membrane array adds 4000 bytes, about 3.9 KiB. Delay queues, synaptic states, input/output buffers and training histories can add more. A useful first acceptance test states an exact input, parameters and expected event times or outputs. 'The plot looks reasonable' is not specific enough to catch a later regression.
CLOSE YOUR NOTES
Can you explain it back?
These questions test the reasoning, not just a remembered phrase. Say your answer out loud or write a sentence first.
01Does an 8-byte edge estimate describe Python object memory?
Reveal the reasoning for question 1
No. It is a declared compact-payload assumption. Python containers have additional overhead that must be measured or accounted for separately.
02When should you scale from three neurons to thousands?
Reveal the reasoning for question 2
After the small model's units, direction, update order and expected outputs are tested, and after a concrete question requires the larger scale.
TAKE THE IDEA WITH YOU
Three things to remember.
- Build a connected sequence of small, inspectable artifacts.
- Preserve numerical references while adding complexity.
- Resource estimates guide measurement; they do not replace benchmarks.
The final part turns your project into a claim another person can inspect, question and attempt to reproduce.
Continue to Part 25 →THE NEXT LAYER
Chapter depth map.
This introduction provides a first model. The following chapters are outlines for deeper future editions; their full treatments have not yet been published.
Build the foundations
24.1 Neurons and synthetic sensing · Outline
- 24.1.1 Project 1: single LIF neuron
- 24.1.2 Project 2: 100-neuron SNN
- 24.1.3 Project 3: motion detector
- 24.1.4 Project 4: looming detector
- 24.1.5 Unit tests, traces and simple baselines
24.2 Sensors and connectivity · Outline
- 24.2.1 Project 5: camera to SNN
- 24.2.2 Project 6: event camera to SNN
- 24.2.3 Project 7: fly-inspired collision avoidance
- 24.2.4 Project 8: connectome graph explorer
- 24.2.5 Dataset provenance and transformation tests
Build research evidence
24.3 Simulation and learning · Outline
- 24.3.1 Project 9: real Drosophila subcircuit simulation
- 24.3.2 Project 10: simulated rover
- 24.3.3 Project 11: simulated drone
- 24.3.4 Project 12: hybrid YOLO plus SNN system
- 24.3.5 Project 13: SNN plus reinforcement learning
24.4 Deployment and original research · Outline
- 24.4.1 Project 14: neuromorphic deployment study
- 24.4.2 Project 15: connectome-derived Physical AI architecture
- 24.4.3 Project 16: original research submission
- 24.4.4 Goal, theory, math, data, code, tests and benchmark template
- 24.4.5 Failure cases, experiments, expected observations and extensions
Read with the original sources.
These readings support the concepts and deeper investigation. The worked numbers and illustrations above are teaching examples.
Author-hosted textbook
Neuronal Dynamics — Gerstner, Kistler, Naud & Paninski (2014) ↗
Use for neuron-model prerequisites and notation; this site provides an original curriculum, not a reproduction of the textbook.
Official engineering documentation
PX4 — official simulation guide ↗
Begin with simulation and versioned interfaces. Simulation success alone does not establish hardware flight readiness.
Official hardware research resource
Intel — neuromorphic computing research ↗
Study event processing, memory and mapping constraints. Do not assume public hardware access or transfer platform-specific energy claims to a new task.