loss as a function of error
— L2: e²
— L1: |e|
hover to compare values
regression: drag points to add outliers
drag data points
— L2 (MSE)
— L1 (MAE)
L2 slope = ·
L1 slope = ·
gradient: how each loss reacts to errors
— L2: 2e
— L1: sign(e)
L2 gradient grows with error — L1 gradient is constant
huber loss: the compromise
— L2
— L1
— Huber
quadratic below δ, linear above
signal fitting: same network, different loss
— true signal
— L2 network
— L1 network
— L2 MSE
— L1 MSE
epoch 0 / 3000