Learning visual representations with projection pursuit