A short intro to error, gradients, and learning rate. Then try the same idea below.
Animated explainer · 75s
🤖
This model has one weight, W. Move it until the blue line fits the dots. The red bars show how much error is left.
The data — look for the rule
Each red bar is one mistake: the gap between the line and a real dot.
real data prediction error
Knob W
3.00
Total Error
30.0
Your Best
—
The error valley — find the low point
Each value of W gives one error. Training moves W toward a lower error.
Knob W3.00
Learning rate — how big each step is0.012
STEP 1
Initialize
STEP 2
Forward
STEP 3
Loss
STEP 4
Derivative
STEP 5
Back-prop
STEP 6
Update W
STEP 7
Repeat
These 7 steps are the training loop. Start with an initial value, make a prediction, measure the error, move the weight in a better direction, and repeat. Larger networks use the same loop with many weights.