On 1/n neural representation and robustness