I am working in the field of object recognition by neural network. I read in some materials about "universal feature extraction" but I don't know what it means exactly. Can you explain it for me? How can we show that the feature is universal or not?
it's a process where user does not have to specify manually what features to detect, but the NN will form them itself - such features that discriminate the dataset the most.
The wording is like "universal feature-extraction", so the extraction is universal, you don't have to specify it.
Dear Mohammad,
Here are some relatively new references of papers that include integration of universal feature extraction in NNs:
Nishio, T., & Nishio, Y. (2008). Periodic pattern formation and its applications in cellular neural networks. Circuits and Systems I: Regular Papers, IEEE Transactions on, 55(9), 2736-2742.
Song, F., Xu, Y., & Liang, Z. (Eds.). (2009). Advanced pattern recognition technologies with applications to biometrics. Medical Information Science Reference.
He, Y., Feng, G., Liu, F., & He, H. (2010, October). Near infrared face recognition based on wavelet transform and 2DPCA. In Intelligent Computing and Integrated Systems (ICISS), 2010 International Conference on (pp. 359-362). IEEE.
Azeem, S. A., & El Meseery, M. (2011, December). Arabic Handwriting Recognition Using Concavity Features and Classifier Fusion. In Machine Learning and Applications and Workshops (ICMLA), 2011 10th International Conference on (Vol. 1, pp. 200-203). IEEE.
Song, Y., Diao, Z., Wang, Y., & Wang, H. (2012, June). Image feature extraction of crop disease. In Electrical & Electronics Engineering (EEESYM), 2012 IEEE Symposium on (pp. 448-451). IEEE.
Best regards,
Yaakov
"How can we show that the feature is universal or not? " Good question.
Universal feature has to be effective in all applications ever existed and yet to come up. Thus, there is no practical way to prove this capacity (eg being universal) through testing; the only way would be to prove a theorem convincingly demonstrating the feature's effectiveness. Fortunately, people don't commonly mean to prove that certain feature is universal; they rather mean the feature is not specific to any particular application (like cell biology, astronomy, nature scene, etc.).
For example, textures are common in many computer vision applications; that would make many texture feature algorithms to be in this category of 'multi-purpose' nature. Should one be involved into analyzing applications of different nature, they may want to access what morphological particularities are typical in all those applications. The algorithms addressing these types are then practical in this scenario.
Again, much people involved into pattern recognition are practitioners, not mathematicians. Proving features to be universal may need appropriate background.
