Deep Learning on Small Tabular Dataset: Using Transfer Learning and Image Classification

Deep Learning is a subset of machine learning inspired by the human brain. It uses multiple layers of representation to extract specific knowledge from raw input. It is best suited for large image or sound-based datasets. Deep learning methods are generally avoided for small datasets because they tend to overfit. Transfer learning can be one approach used to solve this problem. However, in the case of tabular datasets, their heterogeneous nature makes transfer learning algorithms inapplicable. This paper aims to discuss a few approaches using a literature review to convert tabular data into images to overcome such limitations. The paper provides a 2-part study wherein we first give a brief overview of transfer learning enhancing the efficiency of deep learning algorithms and drastically reducing the training time for small datasets. Secondly, we provide a detailed study of different techniques available to convert tabular data into images for image classification such as SuperTML, IGTD, and REFINED approach. Furthermore, we propose a novel approach inspired by IGTD to create a blocked image representation of the tabular data on which we apply transfer learning to demonstrate the application of deep learning methods on small tabular datasets (with less than 1000 data points).