146 karma · joined March 18, 2016
Deep learning models are black box models, which means they are not formally interpretable. You can still sometimes get interpretations out of these models using various methods, but they are fairly underdeveloped, and without actual theories of neural network behavior I don't see that improving anytime soon.
You're right, it's true with or without ML. In fact, it's true with human-run systems, too. Consider a police officer who is more likely to pull people over in a certain neighborhood, and that neighborhood was 95% AfAm. IF they were asked why they pulled over more people in that neighborhood, they could say they were discriminating against the neighborhood, which, in and of itself is not racial discrimination. Of course, further inspection of that police officer's records could show that they are biased towards neighborhoods with a high AfAm population, which would be racial discrimination.
The same scenario can easily arise in an ML context, but interrogating a machine is a very different context from interrogating a human. First of all, people believe that computers are innately 'unbiased' because they are computers, so making the case that an algorithm is biased is already more difficult. Second, going back to the point I made before -- interpreting a model is not the same as providing a human-readable explanation. Asking a question about racial bias in a model which doesn't even encode for race (as many do, in an ill-conceived attempt to be 'neutral') requires skilled people to understand how to ask the question and how to interpret the answer. There's no plug and play process that one can follow to "scrutinize the solution to see if it's complete enough and correct enough".
First of all, deep learning absolutely is used in mission critical systems: https://www.techrepublic.com/article/intel-and-ge-healthcare...
Second, simply because designers use systems that are formally interpretable, that doesn't mean that designers know exactly what causes a given response in the model. Formal interpretability means you know what function the model is approximating -- there's a long way from that to a human-readable interpretation.
Finally, even if the designers can tell you what features cause a certain response in the model, they can only tell you that for features that are encoded in the data! There are plenty of features that aren't encoded in your data that can nevertheless affect the outcome of a model!
The old man the boat.
The complex houses married and single soldiers and their families.
The man whistling tunes pianos.
Time flies like an arrow; fruit flies like a banana.
The cotton clothing is made of grows in Mississippi.
I convinced her children are noisy.
Gets all of 'em wrong.