Article Preview
TopIntroduction
Developing neural network models has played an essential role in many big data fields because big data are the foundation of deep networks. However, because the cost of errors is unacceptable, a big data model cannot be applied in big data application scenarios with high-risk decisions if it cannot provide a warning before the error occurs and cannot provide the interpretation for decision-making. Therefore, people want the model to explain each decision and the credibility of the decision and explanation. Solving these two problems will enhance human trust in the model decisions and allow the model to be used in high-risk big data application domains such as medicine and autonomous driving. This study addresses only one of these questions, and this paper is pioneering in combining them.
This research aims to: 1) Let the model explain the decision; 2) use the uncertainty to provide confidence for the decision to realize the early warning of risk decision.
The interpretable research (Li et al., 2018; Simonyan et al., 2014; Zeiler & Fergus, 2014) on machine learning models of big data is divided into post hoc explanations and self-explanation models. A post hoc explanation is usually questioned because the explanation is not necessarily loyal to the original model (Rudin, 2019). In contrast, a self-explanation model can provide a loyal explanation because the explanation module is embedded in the model (Chen et al., 2019). The self-explanation model has good application in image recognition, especially fine-grained recognition.
Research (Chang et al., 2020; Isobe & Arai, 2017; Kendall et al., 2015; Shi & Jain, 2019) on the decision-making credibility of big data models mainly introduces uncertainty into the model and uses uncertainty to represent the trust degree of the model for its decision-making. Using the trust degree of decision can realize the early warning of risk decision and reduce the probability of error.
The self-explanatory prototype-based big data model explains the model decisions, but does not provide an early warning for risk decisions. Introducing uncertainty into a neural network can help it determine risk predictions, but such a neural network cannot analyze and optimize risk predictions. This paper is pioneering in combining the two. The model will not only provide an early warning of risk predictions, but will also use uncertainty to analyze the reasons for the occurrence of risks and how to optimize the model in the future. The research on the big data model in this paper mainly focuses on the image recognition model because the image recognition model has more research on interpretability, and the explanation of image recognition is easier to understand.
Chen et al. (2019) proposed a prototype-based interpretable fine-grained image recognition network called ProtoPNet. ProtoPNet constructs a model that can generate interpretations by itself according to the way humans think in image recognition (Figure 1). ProtoPNet learns some prototypes (i.e., image blocks representing category features in the training set), which are the basis for classifying images.
However, ProtoPNet cannot make an early warning for possible risk decisions, nor can it perform further explanation analysis for risk prediction to help optimize the model. This paper solves this problem by introducing uncertainty to ProtoPNet.