✳ AN INTERACTIVE MACHINE-LEARNING LABORATORY
Watch a network learn to see.
Draw a little universe. Teach a network its patterns.
Look inside, and see how understanding takes shape.
INITIALIZING INSTRUMENTS/CIRCLES · SEED 42
Two experiments. One starting point.Same data, split, seed, and update count.
03 THE DECISION LANDSCAPE
Where patterns become visible.
EPOCH0000/ 3000
EXPERIMENT A2 → 6 → 4 → 1
EXPERIMENT B
A · HELD OUT—
B · HELD OUT—
CLASS A
0.0 uncertain · 0.5 1.0
CLASS B Click a point to follow its journey through the network. Or draw your own.
Add a point by coordinates ＋
05 THE LEARNING RECORD
Evidence, one update at a time.
TRAINING LOSS—— accuracy
HELD-OUT LOSS—— accuracy
GENERALIZATION GAP—
Train the network to see whether its pattern holds beyond the points it learned from.
Training Held out Experiment BEPOCH →
A model is a hypothesis, drawn in color.The landscape shows predictions between the observations. Its confidence is not a guarantee.
Read the field notes