Alignment with human representations supports robust few-shot learning