Latin hypercube initialization strategy for design space exploration of deep neural network architectures

In recent decades, deep learning approaches have shown impressive results in many applications. However, most of these approaches rely on manually crafted architectures for a specific task in large design space, allowing room for sub-optimal designs, which are more prone to be stuck in local minima and to overfit. Therefore, there is considerable motivation in performing architecture search for solving a specific task. In this work, we propose an initialization technique for design space exploration of deep neural networks architectures based on Latin Hypercube Sampling (LHS). When compared with random initialization using standard datasets in machine learning such as MNIST, and CIFAR-10, the proposed approach shows to be promissory on the neural architectural search domain, outperforming the commonly used random initialization.