Improving self-supervised representation learning by synthesizing challenging negatives - Naver Labs Europe