Two Hundred Photos, One Average

A classifier sorts photos into cat, dog or bird, with exactly one correct animal per photo. It ends in softmax and is scored with categorical cross-entropy: the loss on one photo is

$-\ln(\text{probability given to the true animal})$

where $\ln$ is the natural logarithm, and the validation loss is the average over the validation photos.

On 200 validation photos the loss is $0.35$. A teammate reads that as "the model typically gives the true animal about $70\%$".

Looking photo by photo, the team finds two groups:

• on 190 photos, the model gives the true animal a probability of exactly $0.85$;
• on the other 10 photos, it gives the true animal the same probability as each other.

What probability does the model give the true animal on each of those 10 photos?

Round to 3 decimal places.