Build a network
Stack neurons into layers. Each layer's output feeds into the next. The total parameter count (weights + biases) determines model capacity.
Input
10
→
Layer 1
32
→
Layer 2
16
→
Output
1
Parameter count examples
| Name | Architecture | Parameters |
|---|---|---|
| Tiny | 10 → 16 → 1 | 193 |
| Small | 10 → 32 → 16 → 1 | 897 |
| Medium | 10 → 64 → 32 → 1 | 2,817 |
| Large | 10 → 128 → 64 → 32 → 1 | 11,777 |
| Huge | 10 → 256 → 128 → 64 → 1 | 44,033 |
Loading...
Loading...
Loading...
Think Deeper
Try this:
A network with 10 inputs and layers [256, 128, 64, 1] has 44,000+ parameters but only 2,000 training samples. What happens?
More parameters than samples = overfitting guaranteed. The network memorises training data instead of learning patterns. Rule of thumb: keep parameters well below 10x your training samples.
Cybersecurity tie-in: Security datasets are often small (hundreds of labelled attacks, not millions).
A network with 50,000 parameters trained on 2,000 samples will memorise the training data.
Keep architectures small for tabular security data.