Capsule neural nets for graph objects classification

Abstract

A new way to solve the graph classification problem is addressed. The main method utilized is the application of a capsule neural network on graphs. The results achieved include, firstly, the enhancement of the base algorithm for training a capsule network with the possibility of using graphs as an input (a stage of training for permutation invariants of graph vertices’ transformation matrices is included as well as a memory block for trained matrices), and secondly, a proposition of a training set of labeled graph objects, transformed from the MNIST dataset. This opens a perspective for a better classification of graph objects due to preserving of their structure and transformation invariance between layers.

Review of techniques and models used in optical chemical structure recognition in images and scanned documents

“…5. Since CNNs are sensitive to image rotation, affine transformation, and are unable to capture the spatial relation among the parts, using Capsule Networks may deliver better results [8][9][10]. Capsule Networks replace scalar-output feature detectors with vectoroutput capsules and max-pooling with routing-byagreement which adds position invariance.…”

“…By this time, with the developments in hardware, cloud computing and Deep Neural Networks (DNN), all the OCSR stages started to move completely into Machine Learning models, and delivered exceptional results. Employment of attention-based and contextaware image classification models [6][7][8][9][10] removed the necessity of having separate pre-processing phases like noise removal of image restoration. Modern NLP models [11][12][13][14][15] that are capable of being trained to understand and generate complex-structured sequences replaced the expert-driven rules on molecular structure, bonding and formatting details.…”