Neural Perspective to Jigsaw Puzzle Solving

Understanding, predicting the overall configuration from its constituent parts is an important challenge in Machine Learning and Artificial Intelligence. Agents trained to reason about subparts can then be combined to form complicated reasoning systems. Jigsaw Puzzle solving is an avenue in which the di↵erent components have to be correctly placed in order to create a plausible image from its component parts. We analyze this problem using a neural approach since ultimately it will be easier for neural networks to communicate with each other and such subnetworks will help in advancing the research of neural networks in Computer Vision. We present models based on Multilayer Perceptron, CNNs and LSTMs alongside a Markov Random Field based baseline which works on Energy Based Configurations. We perform experiments on a large scale based on the ImageNet dataset which is orders of magnitude larger than the previous evaluations of algorithms developed for the task.