4 ms·
Just read that one-paragraph summary, but... yeah, duh? I thought it was common knowledge, at least within the ML field, that network biases come from skewed tr
by IceMetalPunk 4y ago
Just read that one-paragraph summary, but... yeah, duh? I thought it was common knowledge, at least within the ML field, that network biases come from skewed training data, not from an issue with the model's architecture?
- PaulHoule 4y agoIt's pretty hard to make a training set which is truly representative of the data. For instance if you are collecting images of what could be a cancerous organ there is going to be a different prior distribution of cancers if you (1) sampled the population at random, (2) sampled people who had symptoms, (3) sampled people that following one set of guidelines for mass screenings, (4) sampled people that following a different set of guidelines. It's quite problematic if people train a model based on one of the above and really use it on a different one. Since cancers are rare compared to non-cancers you might want to enrich the data with more cancers to get a really good representation of cancers, but then that throws the prior distribution out of whack. It would be great to have a way to tune the prior distribution separate from the conditional probabilities as in Bayesian modeling but most ML approaches don't really do that.