BACKPROP
VS. GENETICS

Complex ideas. Clearly explained.

Two ways to train a neural network. A study built to compare them.

Read the paper

Co-author & C# experiment builder · Published research

Two ways to learn.

Backpropagation

Measure the error. Adjust the weights. Try again.

Genetics

Compare a population. Keep stronger candidates. Vary and repeat.

Compare

Select

Vary

Follow the gradient.

Backpropagation propagates output error backward through the network to update its weights. The XOR experiment used a learning rate of 0.2 and momentum of 0.0125.

Network: 2 inputs · 2 hidden nodes · 1 output
Activation: Sigmoid

Search a population.

The genetic method compares candidate networks, selects stronger candidates, and introduces variation. The XOR experiment used a population of 500 and mutation rate of 0.05.

Comparison: 30 trials per method, to 95% accuracy
Scope: XOR results shown here; the paper also studies reptile classification and sin(x).

These are the study’s settings, not optimized defaults for every problem. Results depend on the task, implementation, and chosen parameters. Inspect the experiments

View my original presentation ↗

What the experiments showed.

Published XOR training-time bounds, in milliseconds
Method	Minimum (ms)	Maximum (ms)
Backpropagation	163.9362	647.6394
Genetics	169.9704	1863.1769

XOR · 30 trials per method · Time to 95% accuracy · Observed minimum and maximum.

Genetics had the lower median here. Backpropagation had a narrower observed range.

Inspect methods and exact results

The work behind the work.

Explain the depth.

I had to make the research understandable to a non-technical board without losing what mattered.

Build within constraints.

I worked to make the experiments run on my MSI GS65 laptop.