One neuron: a weighted sum, then an activation function
w₁
w₂
bias b
output g(w₁x₁ + w₂x₂ + b) over the input plane
Backpropagation by hand: 2 inputs, 3 sigmoid hidden units, 1 sigmoid output, L = ½(o − t)², t = 0.85
one step: forward, backward, update
100 steps
reset to the exercise
learning rate η
Train a small network in your browser (full-batch gradient descent, cross-entropy)
hidden units (0 = logistic regression)
learning rate
▶ train
random start
identical start
decision regions and training points
training loss