Semi-supervised active recognition method based on cognitive information particles

Abstract

The invention discloses a semi-supervised active recognition method based on cognitive information particles, including: 1. Using a small data set to train an initial depth model, judging the degree of confidence in the recognition result, and outputting a certain classification result for a certain sample; otherwise, the model will Ask experts to help analyze uncertain samples and give guidance information for uncertain samples; 2. In the model upgrade stage, calculate the cognitive error information of the samples in the expert guidance sample set, and comprehensively consider the cognitive error calculation The cognitive information particle information of the sample, select the sample with a larger value of the cognitive information particle as the target sensitive sample, and determine the number of sensitive samples to be selected; 3. Add the target sensitive sample to the training data set to fine-tune the depth model. Repeated execution gradually improves the recognition accuracy and recognition efficiency of the deep model to adapt to the pattern changes brought about by complex targets and environmental changes. This method introduces cognitive knowledge into the deep learning model, which has profound significance.

Description

technical field [0001] The invention belongs to a realization of cognitive knowledge applied to semi-supervised active recognition in the field of deep learning, and specifically relates to a semi-supervised active recognition method based on cognitive information particles. Background technique [0002] Deep learning is a new field in machine learning research. Its motivation is to establish and simulate the neural network of human brain for analysis and learning. It imitates the mechanism of human brain to explain data, such as images, sounds and texts; deep learning combines Low-level features form more abstract high-level representation attribute categories or features to discover distributed feature representations of data. [0003] Traditional deep learning is a one-way open-loop process, which cannot satisfy human learning methods, such as repeated weighing and comparison to recognize new objects from rough to fine, and from complex to simple. If the deep learning mo...

AI Technical Summary

Problems solved by technology

[0003] Traditional deep learning is a one-way open-loop process, which cannot satisfy human learning methods, such as repeated trade-offs and comparisons to achieve recognition of new objects from rough to fine, from complex to simple

If the deep learning model is trained with more data, it can exhibit the same characteristics as human learning; however, in most practical applications, especially in the industrial or medical fields, the targets to be identified are usually complex or non-uniformly distributed, The target mode also changes a lot, and it is also difficult to collect a large number of labeled samples at one time, while deep learning requires a large number of training samples to train the model to achieve the desired recognition effect

[0004] At present, active learning is mainly based on the size of the uncertain information of the samples to be identified to select samples for labeling, and then use the labeled samples to train the model. This type of technology does not fully consider the cognitive knowledge and cognitive behavior of the model, and is not suitable for deep The practical application of learning; semi-supervised classification is the best solution to the practical application of deep learning, mainly considering how to use a small number of labeled samples and a large number of unlabeled samples for training and classification, so how to choose the most effective samples Improving the performance of deep models as training data to adapt to new application changes is the core issue in the process of deep machine learning. Because this work is very challenging, there is no relevant literature on the active recognition of high-dimensional data based on deep models. or categorized questions

Images