Deep Learning Concept

Backpropagation

Run forward to a loss — then watch the gradient flow backward, multiplying local derivatives.

forward → ← backward (gradients) w0.80 z = w·x+b0.80 a = σ(z)0.69 L = ½(a−y)²0.05 ∂L/∂a −0.31 ∂a/∂z 0.21 ∂z/∂w 1.00 ∂L/∂w = −0.07
prediction a
0.69
loss L
0.048
gradient ∂L/∂w
−0.066
chain rule: multiply the local derivatives right-to-left to get ∂L/∂w
backward.py