Understanding and Mitigating the Impact of Model Compression for Document Image Classification.

Compression of neural networks has become a common norm in industrial settings to reduce the cost of inference and deployment. As document classification is a common process in business workflows, there is a dire need of analyzing the potential of compressed models for the task of document image classification. Surprisingly, no such analysis has been done in the past. Furthermore, once a compressed model is obtained using a particular compression algorithm (which achieves similar accuracy to the uncompressed model), the model is directly deployed without much consideration. However, such compression process results in its own set of challenges which have large implications, especially when the uncompressed model is ensured to be fair or unbiased such as a fair model making critical hiring decisions based on the applicant’s resume. In this paper, we show that current state-of-the-art compression algorithms can be successfully applied for the task of document image classification. We further analyze the impact of model compression on network outputs and highlight the discrepancy that arises during the compression process. Building on recent findings in this direction, we employ a principled approach based on logit-pairing for minimizing this deviation from the functional outputs of the uncompressed model. Our results on Tobacco-3482 dataset show that we can reduce the number of mismatches between the compressed and the uncompressed model drastically in both structured and unstructured cases (by more than 2x in some cases) by employing label-preservation-aware loss functions. These findings shed light on some important considerations before the compressed models are ready for deployment in the real world.