Computational model for neural architecture search

A long-standing goal in Deep Learning (DL) research is to design efficient architectures for a given dataset that are both accurate and computationally inexpensive. At present, designing deep learning architectures for a real-world application requires both human expertise and considerable effort as they are either handcrafted by careful experimentation or modified from a handful of existing models. This method is inefficient as the process of architecture design is highly time-consuming and computationally expensive. The research presents an approach to automate the process of deep learning architecture design through a modeling procedure. In particular, it first introduces a framework that treats the deep learning architecture design problem as a systems architecting problem. The framework provides the ability to utilize novel and intuitive search spaces to find efficient architectures using evolutionary methodologies. Secondly, it uses a parameter sharing approach to speed up the search process and explores its limitations with search space. Lastly, it introduces a multi-objective approach to facilitate architecture design based on hardware constraints that are often associated with real-world deployment. From the modeling perspective, instead of designing and staging explicit algorithms to process images/sentences, the contribution lies in the design of hybrid architectures that use the deep learning literature developed so far. This approach enjoys the benefit of a single problem formulation to perform end-to-end training and architecture design with limited computational resources --Abstract, page iii