Unconventional Wisdom: A New Transfer Learning Approach Applied to Bengali Numeral Classification

In this modern age, natural language processing (NLP) is evolving due to advances in the field of deep learning and its access to huge amount of data and computation power. Recently a lot of attention has been given to OCR for Bangla, the 5^th most widely spoken language in the world. This paper reports on certain rather unconventional transfer learning approaches used to attain 6^th place in the Kaggle Numta competition, where the challenge was to classify images of isolated Bangla numerals. The best result reported in this paper is an accuracy of 97.09% on the NumtaDB Bengali handwritten digit datasets test set, which was obtained by freezing intermediate layers. The unconventional approach used in this paper produces better results than conventional transfer learning while taking less epochs and having almost half the number of trainable narameters.