Taxonomic Bias is Important to Binary Dendritic NN Architecture
Branching dendritic neural network (DNN) architectures with cross branch evaluation for uncertain samples provide a biological analogue to their natural counterpart and allow accuracy of network structures to be improved. In the course of research, eliminating human bias is often a fundamental task to ensure optimally meaningful experimental output. In the case of DNNs, however, human bias through taxonomical reasoning in categorical branching is vital to accuracy.
A convoluted neural network (CNN) was constructed to classify the CIFAR-10 dataset - which is composed of low resolution 32px^2 images of animals and machines. Each type of animal (bird, cat, dog, deer, frog, and horse) and each type of machine (airplane, automobile, ship, and truck) are given unique labels along with their RGB pixel data to allow for training (50K samples) and testing (10K samples) of neural network visual processing capabilities.
Standalone, the basic CNN was able to achieve ~78% accuracy, and when arranged in a taxonomically biased DNN, the same network was able to achieve 85-90% prediction accuracy against the CIFAR-10 test dataset - a significant improvement. The first level of the DNN, animal_vs_machine, was able to achieve ~98% prediction accuracy, for example.
To test the hypothesis that human taxonomical bias may affect the prediction accuracy of the DNN, another framework was implemented where the hierarchical structure is auto-generated. For example, instead of level 1 being animal vs machine, it was generated as labels [0-4] vs [5-9] and trained with the same learning rates. It was found that the auto-generated structure caused overfitting with a large gap of 8-9% developing early causing the level 1 prediction accuracy to plateau at ~82%.
Thus, in DNN architectures, human taxonomic bias has a significant positive impact on prediction accuracy with respect to the CIFAR-10 dataset. This could prove to be valuable to researchers and commercial applications involved in the visual classification of sample images. The perception of the human mind can act as a guide to create parallel reasoning in DNN type architectures with very measurable success.