Paper Title
Deep Input Optimization: Comparison Analysis of Various Optimization Algorithms under Different Input Parametrizations

Abstract
A number of important deep learning (DL) applications, such as adversarial machine learning, single-view reconstruction and AI-driven design are exploiting differentiability of deep neural networks to perform input optimization. In this paper, we address one of the main weaknesses reported in the setup of input optimization algorithms getting stuck in local minima. For our experiments we consider a toy-example of image optimization w.r.t. pretrained MNIST image classification deep neural network, i.e. the task is to find a maximizer of the target MNIST class.As the main contribution of this work, we report comparative results for various optimization algorithms under different input parametrizations used to solve the aforementioned problem. Keywords - Deep Learning, Input Optimization, Autoencoder, Image Classification