High School Lab

Neural Net Internals Explorer

Inspect every weight and activation of a small network — then train it with real gradient descent.

Activation Function

sigmoidrelu

σ(z) = 1 / (1 + e⁻ᶻ) — smooth, output in (0, 1), derivative = a(1−a)

x₁0.000

x₂1.000

w[x1→h1]0.041

w[x1→h2]-0.227

w[x1→h3]0.103

w[x2→h1]-0.028

w[x2→h2]-0.726

w[x2→h3]0.326

b_h1-0.362

b_h2-0.603

b_h3-0.776

v[h1→out]-0.700

v[h2→out]0.099

v[h3→out]-0.175

b_o0.795

Line thickness = |weight|; cyan = positive, red = negative. Circle fill = activation level.

Inputs: x₁ = 0.000, x₂ = 1.000

Hidden Layer

h1: z = -0.3903a = 0.4036

h2: z = -1.3291a = 0.2093

h3: z = -0.4506a = 0.3892

Output

z_o = 0.4652ŷ = 0.6142

For each hidden neuron j (1..3):

z_j = w[x1→hj]·x1 + w[x2→hj]·x2 + b_hj

a_j = act(z_j)

Output:

z_o = v1·a1 + v2·a2 + v3·a3 + b_o

ŷ = act(z_o)