Predict first: if the average error falls, will every example get closer?
| Example | Sleep | Study | Known | Pred. | Error |
|---|
Try a New Input · fixed parameters
What evidence supports this new prediction?
Backpropagation finds gradients · gradient descent updates all 13 parameters
Bias = an adjustable offset added before a neuron's output
Hours ÷ 10 · known scores ÷ 100 · displayed prediction × 100
Cost: mean squared error in scaled score units
Backpropagation computes all gradients at the same state
Gradient descent: parameter − 0.05 × gradient
Weights begin uniformly in [−0.7, 0.7]
Hidden biases: 0.1 · output bias: 0
Knob: negative ← 0 → positive
Angle = atan(weight); ticks are guides, not allowed positions
Train stops at 5000 iterations or a nonfinite result
A good fit to three examples does not establish generalization