What if we've been thinking on Machine Learning and Ethics in the wrong way. A thread.

Machine Learning is, by its very definition the act of learning from data to make predictions, such predictions can be in the way of more data, clustering, actions, etc.

Thus, Machine Learning, by definition, will be prone towards said data biases. Here is where a lot of people start arguing, whether is the job of scientists to debias the data, or the job of the scientists to design algorithms that don't take into account said biases.

One side, like the one where we have @pmddomingos and @ylecun argue that messing with the data defeats the purpose of making algorithms that try to predict how the world is going to act.

The other side where we have @timnitGebru and the now twitterless Anima, argues that we need to take responsibility of our models' results and act accordingly to prevent the models from reinforcing the biases that exist in the data.

Perhaps what we need is a new kind of Machine Learning entirely, the current status quo just doesn't afford to make models that make the world as we think it should be. We just don't have data that is clean enough or analyzed enough.

Perhaps a lot of the effort should be towards just trying and segment the data so that doesn't have as many biases. Like instead of training with a single set for every client, we should train with different sets. Or try and run experiments with different populations.

I think we need to add new steps to ML models, like the ones with Cross Validation, but another like Bias-Validation, to see if our models are biased towards certain populations.

If they are, take responsibility and own the results that we get. If they aren't well, we just can agree to have an unbiased model.

If said analyzes were standard in most papers, we wouldn't be arguing endlessly whether an up-scaler generates white male faces. Everybody would just agree that their Bias Validation goes in that way and be done with it.