Incremental learning method and device applied to intelligent robot

Abstract

The invention discloses an incremental learning method and device applied to an intelligent robot. The method comprises the following steps: in a model initialization stage, inputting a mixed data set into a first deep learning neural network model, and completing initialization of the first deep learning neural network model to obtain an initial weight parameter of the first deep learning neural network model; in a knowledge distillation stage, continuing to train the first deep learning neural network model according to a cross entropy loss function and a distillation loss function to obtain a second deep learning neural network model and a second weight parameter thereof; in the weight alignment stage, alignment adjustment is carried out on the second weight parameter of the second deep learning neural network model according to the initial weight parameter of the first deep learning neural network model, and a third deep learning neural network model and a third weight parameter thereof are obtained. According to the technical scheme, the problem of disastrous forgetting in incremental learning is greatly relieved, and the memory and executive power of the intelligent robot on old tasks are improved.

Description

technical field [0001] The invention relates to the technical field of intelligent robots, in particular to an incremental learning method and device applied to intelligent robots. Background technique [0002] At present, intelligent robots are increasingly widely used in public places such as shopping malls, airports, and stations. Segmentation, detection, and identification of buildings, green belts, pedestrians, and vehicles have become their essential functions. The realization of these functions depends on complex deep neural networks such as ResNet and YOLO and their supporting learning algorithms. Although the structure of deep neural networks is complex and the implementation strategies of learning algorithms are different, they all require huge data sets to ensure that the models are competent for tasks in real scenarios. The production of the data set is mainly divided into three steps: 1) The data collection personnel use a professional camera to shoot the targe...

AI Technical Summary

Problems solved by technology

[0005] (1) The existing deep learning models do not perform well in training on continuous stream data. When the model is trained on new tasks, the performance on old tasks usually drops significantly;

[0006] (2) The camera equipment that collects data before the model is used is different from the camera equipment that acquires data when the model is used, resulting in poor data consistency

Images