Representation learning method based on superimposed convolution sparse auto-encoder

Abstract

The invention discloses a representation learning method based on a superimposed convolution sparse auto-encoder. The method is characterized by comprising the following steps of 1) designing and realizing a reconstruction independent component analysis algorithm including whitening, taking an image data set as input, and iteratively optimizing and learning an output reconstruction matrix thereofto obtain a trained sparse auto-encoding model; 2) constructing a semi-supervised superimposed sparse auto-encoder model to train the feature representation; 3) constructing a convolution model to extract block features from the data, and generating convolution feature representation by applying convolution and pooling operations; 4) superposing the convolution sparse auto-encoder, and further optimizing the convolution feature representation; and 5), training a classifier on the image data set by using a logistic regression model and obtaining a classification result. According to the method,the characteristics of the auto-encoder model and the convolution pooling structure are combined, and a small part of labeled data in a large-scale data set is utilized so that feature representationvectors are optimized, and the classification accuracy of the image data set is improved.

Description

technical field [0001] The invention relates to the field of data mining research, in particular to a representation learning method based on a stacked convolution sparse autoencoder. Background technique [0002] Representation learning is an effective method that can learn higher-level and more robust feature vectors from raw data such as image datasets, and can perform machine learning tasks such as classification and prediction based on the learned feature vectors. In recent years, methods based on deep learning have been widely used in representation learning. According to the usage of labels, they can be mainly divided into supervised deep learning and unsupervised deep learning. A typical supervised deep learning model is a convolutional neural network (CNN), which has been widely used in machine learning tasks such as computer vision. The main advantage of CNN has two aspects. First, its deep structure characteristics greatly increase its learning ability, which can...

AI Technical Summary

Problems solved by technology

The basic idea of ​​SDA is to learn the reconstructed feature representation of the original data through the encoding layer and the decoding layer. Although its goal is to learn high-level feature representation, compared with the convolutional network, SDA discards the inherent structure of the input data. In order to solve For this problem, there have been some works such as stacked convolutional autoencoder and stacked sparse autoencoder trying to combine SDA with convolutional pooling structure. Coexistence of labeled data, ignoring labeled data leads to unsatisfactory performance of unsupervised deep learning models

[0004] Although the above-mentioned supervised and unsupervised deep learning models have achieved good performance in the field of representation learning and have been applied in various fields, there are still two main problems that hinder the further application of these algorithms. The first problem is The training problem of multi-layer convolutional neural network, although some optimization work has been done in recent years, in semi-supervised or unsupervised neural network, some effective regularization and optimization methods, such as sparse constraints, etc., its performance and performance are still Poor; the second problem is the use of label information and the data redundancy of image data. In actual data sets, a small number of labeled data and a large amount of unlabeled data often coexist. Ignoring labeled data will lead to the overall performance of the model. At the same time, because the adjacent pixels of the image in the local area of ​​​​the image dataset are highly correlated, if the data redundancy problem is not solved, the learned high-dimensional feature representation may not be abstract and robust enough, leading to errors

Images