Roman Alexander Englert

The Neural Network Integrity Constraint (NNIC) approach is a novel technique that aims to improve the accuracy of neural network (NN) classifications by reducing misclassified instances using integrity constraints (ICs) derived from a held-out residual dataset (separate from the final test set). This ensures a leakage-free evaluation. The paper describes the NNIC approach and evaluates the effectiveness of the NNIC approach through six publicly available datasets. For each experiment, the dataset is split into different proportions of training and test data to assess the impact of varying data sizes on classification accuracy. The NNs are trained using the training data, and the misclassified instances from the residual data are identified. ICs are subsequently generated using the Random Forest Classifier (RFC) algorithm and applied at inference time to reduce classification errors. The number of test examples (only experiments with ICs) provides with 65,640 an intensive size for the evaluation of misclassified data. The results demonstrate that the NNIC approach significantly reduces misclassification rates across all six experiments, even with halved training data. The experiments record also the number of NN parameters as well as training and IC runtimes. These quantities are indicators of resource consumption but are not direct energy measurements. The NNIC approach relates also to Green AI since a significant reduction in computational effort and error rates is enabled.

Applied Artificial Intelligence (AAI) and, especially Machine Learning (ML), both had recently a breakthrough with high-performant hardware for Deep Learning [1]. Additionally, big companies like Huawei and Google are adapting their product philosophy to AAI and ML [2–4]. Using ML-based systems require always a training data set to achieve a usable, i.e. trained, AAI system. The quality of the training data set determines the quality of the predictions. One important quality factor is that the training data are unbiased. Bias may lead in the worst case to incorrect and unusable predictions. This paper investigates the most important types of bias, namely syntactic and semantic bias. Countermeasures and methods to detect these biases are provided to diminish the deficiencies.