바로가기 메뉴 본문 바로가기 주메뉴 바로가기
  • 10-3bHave you provided an explanation to users about the inference results of the AI model?
    • Indicators like precision, recall, and mean Average Precision must be computed with uncertainty to explain a model’s inference result. Uncertainty is the size of the variance of random variables, and it is an indicator that shows the certainty of the results produced by an AI model. Uncertainty estimations include Bayesian neural networks, ensembles, and dropout.
    ü Dropout is a technique where nodes in a neural network and connections between each node are randomly selected and removed.
    ü Taking advantage of the feature that neural networks implemented with dropout and Bayesian neural networks each generate different neural networks, the same input can be given to the generated neural networks to produce numerous outputs. Then the mean and variance of the outputs can be calculated, and the variance here is the uncertainty.

    • A model’s inference results can be explained by calculating and combining the output performance (e.g. precision, recall) and uncertainty of the AI model. For example, when there is a true or false prediction model, the grounds of “the model’s predictive probability is 98%, which is high, and the uncertainty of the probability is low at 1%, which indicates the result of ‘true’ is reliable” could be provided to the users.

    • However, there must be an explanation about the model’s inference results if the predictive probability is lower than the threshold or if the uncertainty is higher for users to recognize this.

    • The threshold for inference results can be classified as the threshold for performance indicators (e.g. precision, recall) of the AI model and the threshold for the uncertainty of the performance. To generate thresholds for the inference results, define first the possible problematic situations that can occur due to the AI model, and then identify important variables that determine whether a problem has occurred. Problematic situations in this context include not only circumstances that threaten the user’s life or properties but also when the quality is lower than expected or the standard.

    • Your AI model will help you find the threshold. There are various techniques to obtain the threshold of the inference results, and the most common techniques are linear discriminant analysis, support vector machine, convolutional neural networks, and long short-term memory as well as relatively recent techniques such as Graph Extrapolation Networks and simple framework for contrastive learning of visual representations.