With the ever increasing amount of data, the use of more powerful models in data analysis has been expanded. Deep neural networks with a large number of layers, known as deep learning methods, have been very popular. Deep learning is a type of machine learning and artificial intelligence that mimics the human mind to learn a particular subject. This type of learning is one of the most important parts of data science, which includes statistics and forecasting modeling. Deep learning is very useful for collecting, analyzing and interpreting large amounts of data, making this process faster and easier. Deep learning is actually a new approach to the idea of neural networks that has been around for many years and shows up in a new and updated format every few years. Reliability and generalizability are one of the most effective factors influencing the performance of models. Therefore, in addition to high accuracy, the model should also be a proper reliable model. Given that the quality of decisions is related to the quality of predictions, researchers are trying to use the data, knowledge and information from past experiences to be able to model the hidden patterns. Therefore, the predictions provided by the models need to be reasonably accurate. Moreover, one of the topics of concern for managers and experts today is to improve the accuracy of forecasting in order to increase the quality of decisions. Accordingly, improving the generalizability of models in the literature of forecasting and modeling has become one of the most challenging research areas. In this regard, the performance reliability is one of the factors affecting the generalizability of the model that has not been considered in modeling. With the proposed approach deep learning will be able to obtian more reliable results by minimizing fluctuations in accuracy in validation data. In other words, the less the accuracy of performance accuracy fluctuations in validation data, the more reliable the performance stability in inaccessible data or tests can be. In this study, by presenting a new branch of modeling called reliability-based models, we intend to obtain a model with more confidence and generalizability in the deep learning process by minimizing the amount of fluctuations in modeling accuracy, which is a measure of reliability. Therefore, in this study, in order to eva‎luate and validate the proposed approach in deep learning and also to prove its effectiveness in improving the quality of decision making in various fields, the standard data set in the field of healthcare and more specifically cancer diagnosis was reviewed and the results were compared and analyzed.

علي زينل همداني، مهدي خاشعي

مهدي بيجاري، صبا صارمي نيا