Enhanced Deep Network Designs Using Mitochondrial DNA Based Genetic Algorithm and Importance Sampling

Abstract

Machine learning (ML) is playing an increasingly important role in our lives. It has already made huge impact in areas such as cancer diagnosis, precision medicine, self-driving cars, natural disasters predictions, speech recognition, etc. The painstakingly handcrafted feature extractors used in the traditional learning, classification and pattern recognition systems are not scalable for large-sized datasets or adaptable to different classes of problems or domains. Machine learning resurgence in the form of Deep Learning (DL) in the last decade after multiple AI (artificial intelligence) winters and hype cycles is a result of the convergence of advancements in training algorithms, availability of massive data (big data) and innovation in compute resources (GPUs and cloud). If we want to solve more complex problems with machine learning, we need to optimize all three of these areas, i.e., algorithms, dataset and compute. Our dissertation research work presents the original application of nature-inspired idea of mitochondrial DNA (mtDNA) to improve deep learning network design. Additional fine-tuning is provided with Monte Carlo based method called importance sampling (IS). The primary performance indicators for machine learning are model accuracy, loss and training time. The goal of our dissertation is to provide a framework to address all these areas by optimizing network designs (in the form of hyperparameter optimization) and dataset using enhanced Genetic Algorithm (GA) and importance sampling. Algorithms are by far the most important aspect of machine learning. We demonstrate the application of mitochondrial DNA to complement the standard genetic algorithm for architecture optimization of deep Convolution Neural Network (CNN). We use importance sampling to reduce the dataset variance and sample more often from the instances that add greater value from the training outcome perspective. And finally, we leverage massive parallel and distributed processing of GPUs in the cloud to speed up training. Thus, our multi-approach method for enhancing deep learning combines architecture optimization, dataset optimization and the power of the cloud to drive better model accuracy and reduce training time.