Introduction: ECAI 2016 is the best place for European AI scientific achievements. The conference provides researchers with a good opportunity to introduce and listen to the best research results of contemporary artificial intelligence.
Randomized Distribution Feature for Image Classification
Summary: Local image features can be assumed to be drawn from an unknown distribution. For image classification, such features are compared by a histogram-based model or a metric-based model. These local features are used to quantify a set of histograms. The histogram-based model is very convenient and has a vector representation of the image, but the information may be lost during vector quantization. Unlike the histogram-based model, the metric-based model estimates the metrics on the potential distribution of local features to achieve better predictive performance. However, the model requires higher computational costs, and there is no vector representation of the image.
In order to maintain the advantages of these two modes, we propose a (dual) random distribution feature and use the random Fourier feature to represent the potential distribution of local features of each image as a vector feature. We have proved that the convergence of similarity and distance is based on the characteristics of random distribution. The significant advantage of the random distribution feature is that it has a vector representation and can therefore be efficiently computed as a histogram-based model. In addition, it provides strict theoretical guarantees and competitive performance like metric-based models. Compared to the best algorithm of the results, experiments in three real-world data sets show that our proposed method achieves competitive classification accuracy with faster calculation speed. In addition, we demonstrate that the features we propose can use vector-based learning methods, which have long been extensively studied in the traditional machine learning field to deal with problems in distributed learning.
First author
Hongming Shan
Position: Doctor of Computer Science, Fudan University
Research Interests: Machine Learning, Data Mining, Computer Vision, Dimension Reduction, Random Algorithms, etc.
Related academic papers:
·"Learning Linear Representation of Space Partitioning Trees based on Unsupervised Kernel Dimension Reduction" (IEEE Transactions on Cybernetics, 2016)
· "Real-valued Multivariate Dimension Reduction" (Machine Learning and Application 2013)
Via:ECAI 2016
PS : This article was compiled by Lei Feng Network (search "Lei Feng Network" public number) and it was compiled without permission.
Original paper download
