Contradiction neutralization for interpreting multi-layered neural networks

The present paper aims to propose a new method for neutralizing contradictions in neural networks. Neural networks exhibit numerous contradictions in the form of contrasts, differences, and errors, making it extremely challenging to find a compromise between them. In this context, neutralization is introduced not to resolve these contradictions, but to weaken them by transforming them into more manageable and concrete forms. In this paper, contradictions are neutralized or weakened through four neutralization methods: comprehensive, nullified, compressive, and collective. Comprehensive neutralization involves increasing the neutrality of all components in a neural network. Nullified neutralization is employed to weaken contradictions among different computational and optimization procedures. Compressive neutralization aims to simplify multi-layered neural networks while preserving the original internal information as much as possible. Collective neutralization is achieved by considering as many final networks as possible under different conditions, inputs, learning steps, and so on. The proposed method was applied to two data sets, one of which consisted of irregular forms resulting from natural language processing. The experimental results demonstrate that comprehensive neutralization could enhance the neutrality of all components and represent features across a broader range of components, thereby improving generalization. Nullified neutralization enabled a compromise between neutrality maximization and error minimization. Through compressive and collective neutralization of a large number of compressed weights, it became possible to interpret compressed and collective weights. In particular, inputs that were considered relatively unimportant by conventional methods emerged as highly significant. Finally, these results were compared with those obtained in the field of the human-centered approach to provide a clearer understanding of the significance of contradiction resolution, applied to neural networks.