Angelov, Plamen Parvanov and Gu, Xiaowei (2017) MICE : Multi-layer multi-model images classifier ensemble. In: The 3rd IEEE International Conference on Cybernetics :. IEEE, pp. 436-443. ISBN 9781538622018
DL_MNIST_MICE_v1.5.pdf - Accepted Version
Available under License Creative Commons Attribution-NonCommercial.
Download (1MB)
Abstract
In this paper, a new type of fast deep learning (DL) network for handwriting recognition is proposed. In contrast to the existing DL networks the proposed approach has clearly interpretable structure that is entirely data-driven and free from user- or problem-specific assumptions. It is entirely parallelizable and very efficient. First, same fundamental image transformation techniques (rotation and scaling) that are used by other existing DL methods are used to improve the generalization. The commonly used descriptors are then used to extract the global features from the training set and based on them a bank/ensemble of zero order AnYa type fuzzy rule-based (FRB) models is built through the recently introduced Autonomous Learning Multiple Model (ALMMo) method working in parallel. The final decision about the winning class label is made by a committee on the basis of the fuzzy mixture of the trained ALMMo-0 models (where “0” stands for 0 order meaning that the consequent represent a class label, a singleton, not a regression model as in the first order). The training of the proposed MICE system is very efficient and highly parallelizable. It significantly outperforms the best known methods in terms of time and is on par in terms of precision/accuracy. Critically, it offers a high level of interpretability, transparency of the classification model, full repeatability (unlike the methods that use probabilistic elements) of the results. Moreover, it allows an evolving scenario whereby the data is provided in an incremental, online manner and the system structure is being developed in parallel with the classification which opens opportunities for online and real-time applications (on a sample by sample basis). Numerical examples from the well-known handwritten digits recognition problem (MNIST) were used and the results demonstrated the very high repeatable performance after a very short training process which is in addition to the high level of interpretability, transparency.