2 ms·
Why do you think models are neutral mirrors of the test data? That seems unlikely. I thought they were arguing that models can be designed (intentionally or un
by ZeroGravitas 5y ago
Why do you think models are neutral mirrors of the test data? That seems unlikely.
I thought they were arguing that models can be designed (intentionally or unintentionally) to amplify or dampen biased data.
The metaphor is like "overfitting". You can extract too much from the data and make it bad for data not in your test set.
So is an overfitted model a neutral mirror and also one that's not overfitted also a neutral mirror? That seems odd if they put emphasis on different parts of the data. And we generally try to avoid overfitting don't we? Isn't that an algorithmic choice 'biasing' towards data we haven't included in our initial data?