An Incremental Learning Method Based on Diverse Example Sets of Triplets and Gradient Regularization

Abstract

The invention proposes an incremental learning method, a computer and a storage medium based on triplet diverse example sets and gradient regularization, belonging to the field of artificial intelligence. First, get the predicted sample features and real labels, and input the loss function for backpropagation to update the model parameters; second, calculate the prototype representation of the batch data; third, calculate the number of positive samples and negative examples that should be saved for each category The number of samples; secondly, update the number of examples that should be stored in the example set of the existing category; secondly, score the samples in the positive example set example set, and construct the current category example set according to the scores of the samples; secondly, random sampling Obtain the replay sample set, and then perform forward propagation on the replay sample set and the samples in the batch data; secondly, calculate the gradient of the three loss functions; finally, regularize the three different gradients to obtain the final gradient value. Backpropagating updates. The present invention solves the problem of catastrophic forgetting.

Description

An Incremental Learning Approach Based on Triad Diverse Example Sets and Gradient Regularization technical field The present invention relates to a kind of incremental learning method, relate in particular to a kind of based on triplet diverse example set and gradient regularization The improved incremental learning method, computer and storage medium belong to the field of artificial intelligence. Background technique Incremental learning is an important learning and training method in the field of artificial intelligence. In the case of serialized input training data, incremental training updates model parameters rather than training from scratch, and makes The model learns new knowledge without forgetting the old knowledge it has learned. [0003] The example selection strategies adopted by the existing replay-based incremental learning methods are mostly similar to the nearest neighbors. The idea is to select the sample with the closest distance from the center o...

AI Technical Summary

Problems solved by technology

However, the distribution of such an example set cannot well cover the data distribution of the original training samples, so the optimal spatial distribution of the old task cannot be well described in the subsequent replay training.

At the same time, most of the existing incremental learning methods based on replay put the example data into the distillation loss function for optimization. The methods used are relatively simple, and the example data cannot be used efficiently to constrain the direction and size of the optimization.

Images