Input
Layer 1
W₁ (fixed)
Layer 2
W₂ (fixed)
Layer 3
W₃ (fixed)
Output
Weights
Loaded once. Stay in memory.
Activations
Flow through, every inference.