Learning to Walk in the Park

I had this piano teacher who used to make me do the wildest things - play with my arms crossed, in the dark, sometimes even blindfolded. It sounds ridiculous, but the logic was clear: if I could handle those chaotic challenges, then performing under normal conditions would feel like a breeze.

And honestly? It worked. For me, at least. As a human, I got better.

But recently I’ve been training a model - and I noticed something strange. When I feed it complex, improvised versions of a tune, it actually learns some pretty cool harmonic features. It thrives in that chaos.

But when I give it something super simple - just basic chords or a clean melody - it weirdly starts to struggle.

There’s a term for this in machine learning: lack of shared representations.

In plain terms, it means the model never developed a general internal structure that works across both complex and simple input. Like... it knows how to survive in the jungle, but it forgot how to walk in the park.

And that, to me, is fascinating.