Learn the fundamental characteristics of handwritten characters for a necessary neural network calibration

Conventional Deep Neural Network (DNN) learning process built up the DNN architecture to fit the input space, in terms of size, structure and neural shapes. But the calibration for the input data space characterisation is never realised. This paper reports “Less is More” principle and its implementation in the development of DNN. The recent phenomenal success of artificial DNN has brought a popular but false belief: the deeper the network structure and the more training data, the better. We believe “Less is More”, as the characteristic of a problem is more fundamental than the association of statistics only. We view each input sample as a composition of linked pixels in a unique pattern, and the features of such linkage should be characterised. Technically, the necessary feature for each sample is embedded in a pixel-based graph and can be extracted, typically ~50 features per sample pattern. With these 50 characteristic features, we extracted a neater network (size of input, hidden and output layers are 64, 6 and 7 neurons, respectively), and achieved the highest accuracy (~99.9% against all the published DNNs regardless how deep, physical or digital they are) and a higher computing efficiency (~58% less CPU runtime). The developed DNNs bears the features of circuitry which offers better explainability. The successful application of Less is More approach supports and promotes its wider applications; it contributes to DNN methodology. We also anticipate that the developed methodology will eliminate the waste of unnecessary computational efforts such as pruning et al.