Model training method, device, equipment and medium for fusion posterior probability calibration

Abstract

The embodiment of the present invention discloses a model training method, device, equipment and medium that integrates posterior probability calibration. The method includes: first inputting a first preset number of samples into the classification model for prediction, and obtaining the prediction of each sample category and predicted posterior probability; then, according to the predicted posterior probability of each sample, the first preset number of samples are divided into probability intervals, that is, the samples are classified into intervals; in each probability interval, each The number of samples of the predicted category and the number of correctly predicted samples in all samples of each predicted category are obtained to obtain the empirical posterior probability; according to the predicted posterior probability and empirical posterior probability corresponding to each sample, the probability loss value is calculated, and Train a classification model based on the probability loss value. Therefore, the embodiments of the present invention match the predicted posterior probability output by the model with the real confidence, avoiding the situation of "blind self-confidence" in the model.

Description

technical field [0001] The invention relates to the field of neural network models, in particular to a model training method, device, equipment and medium for fusion posterior probability calibration. Background technique [0002] As the depth and width of the network model increase, the prediction accuracy of the network model during the training process gradually increases, and the posterior probability of the network model output, that is, the confidence level also gradually increases. For example, when the network model predicts the category of samples , the posterior probability of the output is mostly above 90%. [0003] However, in the actual application of the network model or when it is verified on the verification set, the confidence of most detection targets is very high, generally exceeding 90%, and the confidence of only a few detection targets falls between 10% and 90%. However, the prediction accuracy rate corresponding to the overall detection target is very...

AI Technical Summary

Problems solved by technology

[0003] However, in the actual application of the network model or when it is verified on the verification set, the confidence of most detection targets is very high, generally exceeding 90%, and the confidence of only a few detection targets falls between 10% and 90%. However, the prediction accuracy rate corresponding to the overall detection target is very low, that is, the network model has "blind self-confidence"

Images