Notes You Can’t Delete
depth
findings/Interpretability
Looking inside

What a loss actually measures

Cross-entropy is conditional entropy, in nats. Rank composers by it, and measure how much a concept really adds.

blue note = 90% redundant

Every “loss” on this site is really a measurement of information. A model’s cross-entropy is the conditional entropy of the next note given its context, in nats — how many bits of surprise are left once the model has done its best. That reframing turns the loss into a ruler for the music itself.

Ranking composers by their unpredictability

Score held-out music by conditional entropy and you get a legible ranking: the more a composer’s next note is determined by what came before, the lower the number. The most chromatic, least tonal writing sits highest; strict tonal writing sits lowest.

1.98
Scriabin — highest entropy, hardest to predict
lowest
strict tonal writing — most predictable

How much does a concept actually add?

The same lens measures a concept’s information content: how many nats it adds that aren’t already implied by its surroundings. For the blue note, the answer is almost none — which is exactly why deleting it changes so little.

+0.023
nats the blue note adds beyond its context
90%
of the blue note is already redundant

The verdict

Reconstructibility is 1 − conditional_entropy. A concept reconstructs precisely to the degree that it’s redundant — carried by the grammar around it. Measuring that redundancy directly is the quantitative core of the whole project: the blue note is 90% implied by its context, so a model that never saw one still expects it.

Researcher notes

  • Units. Cross-entropy loss = H(note | context) in nats/token. A pitch-shuffled floor gives the marginal entropy; the gap is how much structure the model captured.
  • Concept information. The difference between a model’s surprise at a concept’s tokens and the surprise a context-only prediction would carry — small for redundant concepts, large for genuinely novel ones like djent’s register.