data
XOR — four clusters
Circles — one inside a ring
Two moons
Spirals
hidden layers
1
2
units each —
32
training points —
30
learning rate η —
0.032
flipped labels —
15%
weight penalty λ —
off
dropout
off
0.2
0.5
step
train
reset weights
new data
epoch
0
accuracy: training
–
· held-out
–