System and Methodology to Generate Dynamic Real-time Transparent and Bias-free Feature Modules used to Augment Model Learning Process

Features are the backbone of any Machine Learning(ML) system. During the t^th epoch of the model learning process, learnt features are the measurable outcomes of the model for a given input function. One of the biggest areas of research has been to try to impose even the slightest amount of control on what features a model learns. Achieving this in-turn poses various advantages, including re-usability of the features. Some of the state-of-the-art reusability techniques work by training the model weights and freeze at a random point k, on the source system and test the accuracy on the target until the ideal k point gives the highest possible accuracy. However, some of the current challenges remain in trying to control what is being sent to the target model and also in sending bias-free content. Thus, if we can aid this by adding a "bias-free control-factor" to the features being re-used in the model we would truly be able to perform many more tasks efficiently and achieve much higher accuracy along with lower consumption of resources required for training a model, thereby enabling ML towards many more industrial applications. We propose to provide a methodology by which we augment dynamic and bias-free feature modules for a model building process by taking an example in a Natural Language Processing(NLP) context.