The current "best" tools for differential expression analysis, e.g. DESeq2 and EdgeR, incorporate normalization factors accounting for library size and composition. Pointedly, these tools do not "just" normalize prior to analysis, they include the normalization factors as features within the model. If you are looking to utilize a neural network, hierarchical bayesian, or other type of machine learning approach to differential expression, determining the appropriate method of normalization will be a make or break point.