deep neural networks frequently suffer from performance degradation when the
training data is long-tailed because several majority classes dominate the
training, resulting in a biased model. Recent studies have made a great effort
in solving this issue by obtaining good representations