Learned feature representations are biased by complexity, learning order, position, and more