Early CNN layers detect simple edges; each deeper layer combines those into progressively more complex, recognisable shapes.
activation = ReLU(sum(weight * input) + bias)
deeper layers = higher-level features
- Filter nodes/layer — convolutional filters in each layer scanning for a visual pattern.
- Network depth — how many stacked layers the signal passes through, from edges to whole objects.
- Feature input rate — how often new visual features are fed into the network.
- Activation threshold — response strength a filter needs before it's considered to have detected its pattern.
This is the same layered edge-to-object hierarchy that lets a phone camera find faces or a self-driving car recognise a pedestrian in milliseconds.