Abstract
Neural architecture search is the solution adapted to overcome the problem of manually designing the neural network and overcoming time consuming and error prone methods. The technique proposed for learning Neural Architecture Search (NAS) is implemented for the Convolution Neural Network (CNN) where the approach is to train the model by searching the structures in order of increasing complexity. The algorithm chosen is branch and bound for the search space while simultaneously learning the surrogate model. The primary purpose of this project is to understand how we can improvise the search method used in the previous state-of- art methods, like the ones used with reinforcement learning and evolutionary algorithm. The experiment focuses on pruning the redundant networks for the search space considering each “cell” rather than the full CNN. The architecture consists of implementing blocks in incremental size and running the algorithm to improve optimization and reduce search time for every network that will train and this process is further refined by using a surrogate function which is the Recurrent Neural Network to train the child models and bring down the time to negligible as compared to the earlier ones. The dataset used for my trial is CIFAR- 10 (collection of images). The method is comparatively simpler and faster in terms of computational speed and number of models evaluated. The model achieves 85.80% accuracy.