Supplementary Materials of ClimbQ: Class Imbalanced Quantization Enabling Robustness on Efficient Inferences Ting-An Chen 1,2, De-Nian Y ang 2,3 Ming-Syan Chen 1,3 1

After the determination of scaled class distributions in Sec. A.2 Appropriate class data size estimation Therefore, we propose a distribution scaling on class variances in Sec. To ensure the homogeneity of the variances, we examine it in accordance with Levene's hypothesis in Sec. Moreover, we derive from the analytical results that the homogeneity criterion is satisfied if the data size of each class is restricted. Given the definitions and the notations in Eq. (3) and its subsequent paragraphs of the The HomoV ar loss is proposed in Sec.