Neural network model training method and device

Abstract

The invention relates to a training method of a neural network model for image recognition, and the method comprises the steps: in a first training stage, calculating a first loss function corresponding to a first group of input image samples, and training the neural network model through the first loss function, wherein the calculation of the first loss function comprises: (1) for each input image sample in the first set of input image samples, calculating a classification item corresponding to a difference between a predicted classification label and a real classification label of the input image sample; and (2) for each input image sample in the first set of input image samples, calculating a reconstruction error term corresponding to a difference between the input image sample and its reconstructed image sample.

Description

technical field [0001] The present disclosure relates to continuous learning scenarios in the field of artificial intelligence. More specifically, the present disclosure relates to a training method and device for a neural network model for image recognition. Background technique [0002] Traditional machine learning is performed for a fixed task, that is, the dataset used to train the learning model contains training data with a fixed distribution. When a new data set (ie, a data set containing training data with a new distribution different from the fixed distribution) is input, the learning model generally needs to be retrained. After retraining, the learning model can only respond to the new data set, but cannot respond to the original data set (ie, the data set containing the data of the fixed category). This problem is called "catastrophic forgetting (Catastrophic Forgetting)" in machine learning. In fact, the "catastrophic forgetting" is the result of the "Stabilit...

AI Technical Summary

Problems solved by technology

However, for the weight regularization method, the input data of previous tasks cannot be obtained; for the expression regularization method, although a small number of input data samples belonging to the previous task can be obtained, sufficient features cannot be extracted from a small number of input data samples learning for new tasks

Therefore, the inability to comprehensively extract features to meet the needs of learning for new tasks is one of the limitations of current category incremental learning.

Images