Industrial Automatic Control Systems and Controllers Annotation << Back Analysis of the Efficiency of Learning Algorithms Based on Gradient Descent in Case of A Pretrained Neural Network Fine-tuning K.A. Barinov, K.V. Mikolaichuk Deep learning algorithms based on gradient descent currently have a success in various computer vision tasks, such as image classification, object detection and semantic segmentation. However, there is a problem of efficient using training data in case of exploiting a trained neural network: big amount of new data, which contain useful features, could increase performance of a neural network if they were included in the training process. In this work we brief an idea of Semi-Supervised learning for object recognition and make a review and comparison of 3 versions of the Pseudo-labelling method. Also the results of experiments show, that Semi-Supervised learning can help to achieve better accuracy (mAP) in task of the MADI History Museum’s exhibits recognition using 2-times less amount of labelled data. Keywords: neural networks, object recognition, Semi-Supervised learning, Pseudo-labelling. DOI: 10.25791/asu.3.2022.1352 Pp. 09-17.