>>12887925A image classification model doesn't learn anything beyond its training set. If it is misclassifying, that could be for several reasons, one of them being that the training set is unbalanced. For example, If your training set for the "human" category has 90% white faces and 10% black faced, the model will tune more towards recognizing the white faces since it can achieve 90% accuracy even if it gets all the black faces wrong.
The Complain that AI research is biased is based on that most training datasets are biased.
The fact that a model misclassifies blacks with gorillas doesn't mean anything deep and insightful about AI that validates the your racist views. It means the model is dumb and making mistakes. I can purposefully create a model or a training set that will mistake whites with zebras, that wont mean anything either.
t. ml researcher
