Non-adversarial generative self-coding method and system based on dynamic curved surface segmentation

Abstract

The invention discloses a non-adversarial generation self-coding method and system based on dynamic curved surface segmentation, and relates to the technical field of data generation, and the method comprises the steps: obtaining a training data set; reducing the dimension of the training data set to a low-dimensional space by using an encoder to obtain a first low-dimensional vector set; dividing a unit sphere in the low-dimensional space into a plurality of regions by using central Willon inner segmentation; determining an area to which each low-dimensional vector in the first low-dimensional vector set belongs by using a minimum loss matching algorithm; calculating a first distribution distance one by one; optimizing the encoder by taking minimization of the first distribution distance as a target; reducing the dimension of the training data set to a low-dimensional space by using the optimized encoder to obtain a second low-dimensional vector set; training and optimizing the decoder by using the second low-dimensional vector set to obtain an optimized decoder; obtaining existing data; and inputting the existing data into the optimized decoder for decoding to generate new data. According to the invention, the data generation quality of non-adversarial self-coding can be improved.

Description

technical field [0001] The invention relates to the technical field of data generation, in particular to a non-adversarial generative autoencoding method and system based on dynamic surface segmentation. Background technique [0002] Current big data algorithms rely on massive amounts of data. But for certain problems, data acquisition can sometimes be very expensive. Therefore, generating new data based on existing data has become a great demand in the industry. In reality, the collection and labeling of data often requires a lot of cost, and some specific data are difficult to obtain. Therefore, generating new data from original data has great demand and application in industry. [0003] Existing data generation schemes are divided into two technical routes: adversarial and non-adversarial. Among them, adversarial generated self-encoded data is of high quality, but it is not stable enough, difficult to train, and requires a lot of parameter tuning. In addition, since th...

AI Technical Summary

Problems solved by technology

In reality, the collection and labeling of data often requires a lot of cost, and some specific data are difficult to obtain, so generating new data from original data has great demand and application in industry

[0003] The existing data generation schemes are divided into two technical routes: adversarial and non-adversarial. Among them, adversarial generated self-encoded data has high quality, but is not stable enough, difficult to train, and requires a lot of parameter adjustment.

In addition, since the anti-autoencoder uses a black box to calculate the distribution distance, it is difficult to explain what information the model remembers and what information it forgets

Therefore, although the anti-self-encoding effect is good, it is inconvenient to use

Non-adversarial self-encoding can solve the problem of inconvenient use of adversarial self-encoding. Although non-adversarial self-encoding is easy to use, the effect is not good. Non-adversarial self-encoding cannot use excessive parameters to calculate the distance between distributions, resulting in distance calculation. inaccurate, resulting in lower quality data generation

Images