Alignment with human representations supports robust few-shot learning | Read Paper on Bytez