Theory 54/Precision & Information/Machine learning
Interpretability Follows Compression
The claim
A model is interpretable when the ratio of its description length to its input dimension is small, so a large model can be more interpretable than a small one.
01Mechanism
Interpretation requires a compact account; a model whose behaviour compresses well admits one, regardless of parameter count.
02Prediction
Interpretability should track the ratio.
03Refutation
Interpretability tracking size would remove the compression term.