Towards Deep Machine Reasoning: a Prototype-based Deep Neural Network with Decision Tree Inference

In this paper we introduce the DMR -- a prototype-based method and network\narchitecture for deep learning which is using a decision tree (DT)-based\ninference and synthetic data to balance the classes. It builds upon the\nrecently introduced xDNN method addressing more complex multi-class problems,\nspecifically when classes are highly imbalanced. DMR moves away from a direct\ndecision based on all classes towards a layered DT of pair-wise class\ncomparisons. In addition, it forces the prototypes to be balanced between\nclasses regardless of possible class imbalances of the training data. It has\ntwo novel mechanisms, namely i) using a DT to determine the winning class\nlabel, and ii) balancing the classes by synthesizing data around the prototypes\ndetermined from the available training data. As a result, we improved\nsignificantly the performance of the resulting fully explainable DNN as\nevidenced by the best reported result on the well know benchmark problem\nCaltech-101 surpassing our own recently published "world record". Furthermore,\nwe also achieved another "world record" for another very hard benchmark\nproblem, namely Caltech-256 as well as surpassed the results of other\napproaches on Faces-1999 problem. In summary, we propose a new approach\nspecifically advantageous for imbalanced multi-class problems that achieved two\nworld records on well known hard benchmark problems and the best result on\nanother problem in terms of accuracy. Moreover, DMR offers full explainability,\ndoes not require GPUs and can continue to learn from new data by adding new\nprototypes preserving the previous ones but not requiring full retraining.\n