Loss Curves in LineRider

Josef Dean: "Sure matplotlib is cool, but what if I want to load my loss curves into the 2006 hit Flash game LineRider?"

This is the kind of thing that sounds like a joke and then you watch the video and it works and you sit there for a moment wondering whether this person has achieved something important.

The confluence of interests required to even conceive of this — you need to know enough ML to care about loss curves, enough nostalgia to remember LineRider, and enough engineering chops to connect the two — is a very specific intersection on the Venn diagram of human experience.

But there's something genuinely beautiful about watching a training run play out as a sledding track. The sharp initial descent. The bumpy middle epochs where the rider catches air. The smooth convergence at the end where the sled glides to a stop — or doesn't, if your model is diverging, in which case the rider launches off a cliff.

Your loss curve as terrain. Your hyperparameters as physics. Your overfitting as a loop-de-loop that sends a tiny pixelated person into the void.

Matplotlib could never.