This is an important topic but the framing is not technically robust. Are convolutions "biased" because they are amplifying variance between adjacent pixels? Y

This is an important topic but the framing is not technically robust. Are convolutions "biased" because they are amplifying variance between adjacent pixels? Yes if your definition of biased is statistical, No if your definition of bias is social. I have $300 for someone who can make a convincing technical argument that the inception architecture [1] is socially biased, without actually training it on a dataset. Until then its the data or implementations that are biased. Maybe your trained model is biased - yep that's the data. Maybe you're cherry picking your algorithm to give you the best result on your chosen dataset - that's a biased implementation. More broadly though, the concept that "systems/people will make inferences on new data which align with patterns they see more frequently" is just a reality of intelligence at this point. Humans constantly fail at this. So yes we should be better and bias is real, to solve it we need to get way more technically accurate on what that bias consists of please. [1]https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/43022.pdf https://static.googleusercontent.com/media/research.google.c...