⚔️ Adam vs RMSprop Visualizer
Same landscape, same starting point — watch momentum change everything.
📋 How To Use This Lab
Press
Next Step ▶
— both balls take one REAL optimizer step, computed live.
Watch the
red RMSprop trail
zigzag as it reacts only to the current gradient's size.
Watch the
green Adam trail
glide smoothly — its momentum (β1) remembers where it was heading.
Use
Auto Run
to race them all the way down and see who actually gets closer to the bottom.
Step 0
RMSprop loss = 2.004
Adam loss = 2.004
The Ravine — Two Valleys Joined by a Flat Saddle
◀ Previous Step
Next Step ▶
Auto Run ▶▶
↺ Reset
🔴 RMSprop — divides by past squared gradients
v = β2·v + (1-β2)·g²
x -= lr · g / (√v + ε)
🟢 Adam — ALSO adds momentum on top
m = β1·m + (1-β1)·g
← the extra step
v = β2·v + (1-β2)·g²
x -= lr · m̂ / (√v̂ + ε)
⚔️ RIVAL
Press Next Step — I'll move both balls using their real, actual update rules.
🧑🏫 PROF. TORCH
Adam is basically RMSprop plus one extra idea: remember which way you were already moving.