What you will learn
Write an application decision sheet that can conclude an SNN is not justified.
Before you begin: Robotics: estimation, planning & control; Neuromorphic computing
START WITH A QUESTION
Picture this.
A model can be elegant and still be the wrong tool. A good use-case decision begins with the job: what must the system detect or control, how quickly, and within which resource budget?
- Baseline
- A reference method against which a proposed change is evaluated.
- Operating requirement
- A measurable condition the system must satisfy in its intended setting.
- Whole-system budget
- The relevant resource use of all included components together.
Read for the idea first. Work through the numbers next. Try answering the questions before opening the explanations.
Write the operating requirement first
A warehouse robot, a wearable sensor and a remote rover face different data, deadlines and failure costs. Describe the input, required output, environment and acceptable error before choosing a neural architecture. For example, detecting brief motion changes may make event sensing interesting; inspecting a static printed label may need information that a quiet event stream does not provide. These are application hypotheses to test, not claims of proven deployments.
Give the simplest baseline a fair chance
A threshold, temporal filter or conventional vision model may already satisfy the requirement. Build that reference before introducing a more complex neural pipeline. Then name the proposed advantage: lower whole-system energy, shorter response time, better robustness or less training data. Match the quality target and include integration costs. A use-case sheet should be allowed to conclude that the new component adds complexity without enough benefit.
Find the dominant part of the budget
Suppose a hypothetical system consumes 8 W in motion and 2 W in computation. Halving computation saves 1 W, reducing the total from 10 W to 9 W, not to 5 W. The full-system improvement is 10%. This is why a large percentage improvement in a small component can disappoint at the battery. For a stationary sensor, computation might occupy a larger share; the same technical change could matter more. Draw the component budget rather than assuming the bottleneck.
Turn energy into an operating estimate
Watt-hours describe energy. Divide usable watt-hours by representative average watts to estimate hours of operation. A 20 Wh budget at 5 W gives four hours under the stated idealization. Sensors, communication and host processing belong in the boundary if they use that energy. Workload variation, conversion losses and battery behavior can change the result. Use this estimate to decide what to measure next, not as a guarantee of runtime.
SEE THE RELATIONSHIPS
From structure to behavior.
MAKE THE IDEA PRECISE
The mathematics.
| Symbol | Meaning |
|---|---|
| usable_energy_Wh | Available energy within the chosen operating assumptions, watt-hours |
| average_power_W | Average whole-system power, watts |
| runtime_hours | Idealized operating duration, hours |
Build the equation, step by step
- Use energy actually available to the system, not just a nominal battery label. Keep the energy and power units compatible.
- Divide watt-hours by watts to obtain hours. The calculation assumes representative average power over the discharge interval.
- Include sensors, communication, processing and actuators in the boundary relevant to the task. Power changes with workload and operating conditions, so the quotient is an estimate rather than a guaranteed runtime.
WORK THROUGH THE NUMBERS
Worked example
An illustrative usable energy budget of 20 Wh and average consumption of 5 W gives 4 hours. Conversion loss, battery behavior and changing workload can reduce real runtime.
A small Python experiment.
Python 3 · Standard library only · Synthetic educational example. Run this in your own notebook or a local file.
usable_wh, average_w = 20.0, 5.0
print(f"idealized runtime = {usable_wh/average_w:.1f} h")Read the code as a story
The quotient uses Wh divided by W, leaving hours. It assumes positive average power and a representative workload. If you change the system boundary by adding a host, change the denominator too. A number without its boundary is not a useful runtime claim.
Expected output
idealized runtime = 4.0 h
Change one quantity, predict the result, and then run the code again. Explain any difference before adding complexity.
Test your understanding.
Try it: Add 3 W of host and sensor overhead. How does the estimated operating duration change?
Self-check: Would an efficient neural chip necessarily extend a robot’s runtime substantially?
Not if motors or other components dominate consumption. Quantify the component budget first. A large reduction in a small component can produce only a small whole-system gain; extra sensors or host computation can cancel it.
Where this model stops: The atlas is a decision framework, not evidence that every listed application benefits from spiking computation.
A research question: Which task-specific constraint would justify paying the extra integration cost?
Record evidence in your mastery worksheet →PREDICT / TRY / EXPLAIN
Make a one-page go/no-go sheet
Use paper, a calculator or the Python example. Write your prediction before checking the result.
- Choose one nonclinical simulated use case. Write the task, deadline, quality metric and simplest baseline.
- Add 3 W of sensor and host overhead to the original 5 W budget. Recalculate duration with 20 Wh.
- State a result that would make you reject the more complex neural design, even if its chip-only power were lower.
Compare with a worked solution
The total becomes 8 W and the estimate becomes 2.5 hours. A reasonable rejection rule might be failure to meet the task deadline or no whole-system energy improvement at matched quality. Choose the rule before the benchmark; otherwise it is easy to keep moving the goalposts.
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.
01How much total power is saved by reducing a 2 W component to 1 W in a 10 W system?
Reveal the reasoning for question 1
1 W, or 10% of the original system total. It is a 50% component reduction, not a 50% system reduction.
02Why include the baseline if the goal is to study a new model?
Reveal the reasoning for question 2
It tells you whether the new mechanism earns its complexity for the chosen task. A negative comparison can still clarify where the idea does or does not help.
TAKE THE IDEA WITH YOU
Three things to remember.
- Start with measurable requirements, not a preferred technology.
- Find the system's dominant costs before optimizing a component.
- Allow a well-designed comparison to reject unnecessary complexity.
Next we turn the book's small experiments into a coherent project with an inspectable result at every stage.
Continue to Part 24 →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.
Machines in physical environments
23.1 Mobile platforms · Outline
- 23.1.1 Micro-drones and autonomous drones
- 23.1.2 Rovers, planetary rovers and autonomous vehicles
- 23.1.3 Humanoids and search-and-rescue robots
- 23.1.4 Agricultural, underwater and space robotics
- 23.1.5 Navigation, collision avoidance and sensor constraints
23.2 Work and infrastructure · Outline
- 23.2.1 Robot arms, warehouse robots and factory robots
- 23.2.2 Motion detection and industrial monitoring
- 23.2.3 Event-based surveillance and smart cameras
- 23.2.4 Swarm robotics and distributed sensing
- 23.2.5 Privacy, environmental variation and deployment constraints
Edge and human-facing applications
23.3 Low-power and adaptive devices · Outline
- 23.3.1 Low-power edge AI and always-on sensing
- 23.3.2 Wearables and adaptive sensors
- 23.3.3 Prosthetics and brain-machine interfaces
- 23.3.4 Clinical translation is a separate validated process
- 23.3.5 Conventional alternatives before biological analogies
23.4 The use-case decision sheet · Outline
- 23.4.1 Problem, sensors, architecture and models
- 23.4.2 Hardware, software and operating environment
- 23.4.3 Derive latency and power budgets from task needs
- 23.4.4 Limitations, fallback and conventional baseline
- 23.4.5 Explicit SNN/connectome justification or rejection
Read with the original sources.
These readings support the concepts and deeper investigation. The worked numbers and illustrations above are teaching examples.
Author manuscript / survey
Gallego et al.: Event-based Vision, A Survey ↗
A field map for sensing, representations and algorithms. Follow its cited primary experiments before asserting an application advantage.
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.