Method, electronic device and storage medium for face recognition network training

Abstract

Embodiments of the present invention relate to the field of deep learning, and disclose a face recognition network training method, electronic device and storage medium. The method, electronic device and storage medium for face recognition network training in the present invention include: searching for network parameters corresponding to at least two convergence points of a pre-trained initial deep learning network; fusing the initial deep learning network at at least two convergence points The corresponding network parameters are generated to generate optimized network parameters; the initial network parameters of the initial deep learning network are changed to optimized network parameters to obtain a new deep learning network. With this embodiment, the performance of the network to solve practical problems can be deeply learned.

Description

technical field [0001] Embodiments of the present invention relate to the field of deep learning, and in particular, to a method, electronic device and storage medium for face recognition network training. Background technique [0002] Accurate deep learning networks rely on good training methods, and often conventional approaches to deep learning networks require first collecting or collecting problem-specific data. Often, the data used to train the network is only a subset of the data involved in a specific problem. After obtaining the data set, design performance indicators, use gradient descent method to iteratively update network parameters, and finally obtain a trained deep learning network. [0003] At present, the deep learning network usually adopts the gradient iterative training method, which essentially performs a series of mathematical operations and converges when the loss function value reaches the minimum value. However, in the process of iterative optimizat...

AI Technical Summary

Problems solved by technology

[0003] At present, the deep learning network usually adopts the training method of gradient iteration, which essentially performs a series of mathematical operations and converges when the loss function value reaches the minimum. However, in the process of iteratively optimizing the network, saddle points and local minimums will be encountered. and other problems, resulting in the inability to train the deep learning network to the optimal state; in addition, because the data set is not completely equal to the data distribution of the actual problem, even if the performance of the deep learning network is optimized on the training data, the depth cannot be guaranteed. Learn the performance of networks in solving real-world problems

Images