Saptarshi Bej

Machine learning (ML) and deep learning techniques can guide therapeutic decision making by learning pattern from existing patient data. However, to train such models reliably we need a large pool of data, which is often unavailable in light of the patient numbers, effort and the costs to generate such data. The lack of training data does not allow us to reliably validate the model which makes it unsafe to be used practically in clinics . We developed a new method of data augmentation called Random Affine Combination Shadow-sampling (RACoS). From a small amount of data, we can create a large number of RACoS samples that capture the pattern present in the data. We then train our machine learning models with the RACoS data. We observed from our initial studies that our approach provides the ML models with a better experience of the data, compared to training the models with small amounts of data. We also used RCoS sampling to quantify the overfitting tendency of the model even if there is not enough data for validation. Since the RACoS samples provide the ML models with a better experience of the data, we are now developing a new learning methodology, based on one-shot learning.