Ethical Evaluation and Optimization of Artificial Intelligence Algorithms Based on Self Supervised Learning

Abstract

Active learning solves the problem of requiring a large amount of manpower and resources due to the large size of training samples. The core problem is how to select valuable samples to reduce annotation costs. Using neural networks as classifiers, most methods choose samples with large amounts of information without considering the issue of information redundancy between the selected samples. Through the study of redundancy issues, the author proposes a sample selection optimization method to reduce information redundancy. Using uncertainty methods to select samples with high information content to form a candidate sample set, and using latent variable vectors calculated in the network to represent sample information, the cosine distance between candidate samples is calculated using this vector to select subsets with large interval distance and low information redundancy. Compared with several uncertainty methods in the Mnist, Fashion mnist, and Cifar-10 datasets, this method reduces labeled samples by a minimum of 11% with the same sample accuracy. The higher the dimensionality of the feature vector calculated by CNN, the more candidate samples it contains, and the more information it contains. After being improved by self supervised learning algorithms, the effect becomes more significant. The more candidate samples are selected, the stronger the information redundancy. The better the performance of self supervised learning algorithms.