Accelerating Machine Learning Research Using Transfer Learning

Abstract

Research and development in machine learning involves a continuous cycle of training, testing and tuning. This is generally a time and computation intensive process, and hence leads to a slow experiment cycle. We intended to accelerate this cycle by using the transfer learning technique. Using transfer learning we can produce the same results in a matter of minutes, which previously took weeks to generate. This technique is also highly beneficial when the training data is scarce and computing infrastructure is limited. Transfer learning has been applied to many fields ranging from image classification, medical diagnosis of tumors, diabetic retinopathy, generating captions, self-driving cars, etc. Transfer learning is considered to be the next driver of machine learning success. In this paper, we introduce transfer learning, describe its implementation, and talk about its various applications and benefits.