Neura saysThe learning rate sets step size — the single most important hyperparameter.
The learning rate scales each gradient step. Too high and training overshoots or diverges; too low and it crawls or gets stuck. There's a sweet spot, and it changes during training — hence schedules (warmup, decay) and adaptive optimizers. If a model won't train, the learning rate is the first dial to check. Get it wrong and nothing else matters.
Power-ups you unlock
Scales each gradient-descent step
Too high → diverge; too low → crawl/stuck
Schedules (warmup, decay) help
First thing to tune when training fails
The Overfit Ogre attacks — common mistakes
Leaving it at a default for every problem
No warmup/decay on large models
Blaming the architecture before the learning rate
Boss battleDescribe the symptoms of a learning rate that’s too high vs too low.
Example code
<!doctype html><html><head><meta charset="utf-8"></head>
<body style="background:#06040d;color:#e6e0ff;font-family:monospace;padding:20px"><pre>too high → loss explodes / bounces
too low → loss barely moves
just right → smooth, steady descent</pre></body></html>