Umelá inteligencia

Pattern recognition is a common task in machine learning, and the deep neural network is a successful model that delivers high performance in pattern recognition [1,2]. A pattern recognition problem entails the mapping of the input data set  to the classification set . Using machine learning, the designer decides which type of classifier and method for parameter identification will be used. From the user’s point of view, the system is a black box, and they receive no explanation as to why a specific result of the classification was selected. The user cannot take responsibility for a decision if they do not understand how the decision was taken. This is not right, especially if the classification is incorrect. Recently, this aspect was discussed in the literature [3–7]. Even the meaning of the term “explainability” itself is under discussion [8,9]. According to [10], “explainability” is closely related to “interpretability”, the ability of a human to understand the classification through introspection or explanation. If the classification result is not explainable, it not only means that the user cannot explain how the recognition was obtained, but they also may not be aware whether they use the pattern recognition system in the proper way, for instance, whether the assumptions under which the pattern recogniser was designed are being met.

From this perspective, the sample novelty means applying the recogniser to samples that are not similar to the samples from the training set. The user may know the pattern classes, but they may not be familiar with the training set. In fact, the recogniser trained by machine learning provides results under the condition that the query is similar to the samples of the training set. For instance, we test the case in which the training set is a clean MNIST database [11], but the test set is the same set with Gaussian noise perturbations [10,12], or the Fashion-MNIST is applied as the test set [13]. The explainable recognition system has to warn the user if they try to misuse it. The lack of training data for remote sensing has been analysed in [14] or in [15] for face recognition.

A similar approach to this is an outliers detection [16,17]. Let us suppose that we are interested in queries that are not present in the training set classes. In this case, we would not be looking for black swans [18], the outliers belonging to the classes within the training set, as the outliers have been defined in [17]. This problem is also known as a one-class classification [19–21], anomaly detection [22], or novelty detection [23]. To bypass the non-existence of training samples outside the training set, one can generate artificial samples. This approach is used in the generative adversarial networks (GAN) [24] and adversarial autoencoders (AAE) [25].

There are three main differences between the GAN/AAE and our approach. First, GAN/AAE approach plays with probability distributions, while our approach only takes into account the average value. Second, GAN/AAE does not take into account the used feature extractor and the classifier in the used pattern recogniser (this can be eliminated by using conditional GAN [26]). Lastly, our goal is to protect the recogniser against unknown samples, and due to the curse of dimensionality a cardinality of unknown samples set is much more higher comparing with the training set. Therefore we prefer to reduce dimensionality using linear methods and afterwards to reduce the training space cover by the nonlinear transform represented by neural network. The question we are asking is whether the training GAN in fake samples will help in outliers’ recognition. This problem also applies to other generative models [27,28].

 

References

[1]    J. Rigelsford, Pattern Recognition: Concepts, Methods and Applications, Assem. Autom. (2014). doi:10.1108/aa.2002.03322dae.002.

[2]    J. O’Rourke, G.T. Toussaint, Pattern recognition, in: Handb. Discret. Comput. Geom. Third Ed., 2017. doi:10.1201/9781315119601.

[3]    F.K. Dosilovic, M. Brcic, N. Hlupic, Explainable artificial intelligence: A survey, in: 2018 41st Int. Conv. Inf. Commun. Technol. Electron. Microelectron. MIPRO 2018 – Proc., 2018. doi:10.23919/MIPRO.2018.8400040.

[4]    A. Adadi, M. Berrada, Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI), IEEE Access. (2018). doi:10.1109/ACCESS.2018.2870052.

[5]    T.W. Kim, Explainable artificial intelligence (XAI), the goodness criteria and the grasp-ability test, (2018). http://arxiv.org/abs/1810.09598.

[6]    O. Biran, C. Cotton, Explanation and Justification in Machine Learning: A Survey, IJCAI Work. Explain. Artif. Intell. (2017). doi:10.1108/13563281111156853.

[7]    J.M. Alonso, C. Castiello, C. Mencar, A bibliometric analysis of the explainable artificial intelligence research field, in: Commun. Comput. Inf. Sci., 2018. doi:10.1007/978-3-319-91473-2_1.

[8]    H. Hagras, Toward Human-Understandable, Explainable AI, Computer (Long. Beach. Calif). (2018). doi:10.1109/MC.2018.3620965.

[9]    B. Mittelstadt, C. Russell, S. Wachter, Explaining Explanations in AI, (2018). doi:10.1145/3287560.3287574.

[10]  N. Akhtar, A. Mian, Threat of Adversarial Attacks on Deep Learning in Computer Vision: A Survey, IEEE Access. 6 (2018) 14410–14430. doi:10.1109/ACCESS.2018.2807385.

[11]  Y. Lecun, THE MNIST DATABASE of handwritten digits, Html(MNIST Dataset). (2010).

[12]  S. Dube, High Dimensional Spaces, Deep Learning and Adversarial Examples, (2018). http://arxiv.org/abs/1801.00634.

[13]  X. Han, R. Kashif, R. Vollgraf, Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms, Arxiv. (2017). https://arxiv.org/pdf/1708.07747.pdf (accessed September 15, 2017).

[14]  D. Tuia, C. Persello, L. Bruzzone, Domain adaptation for the classification of remote sensing data: An overview of recent advances, IEEE Geosci. Remote Sens. Mag. 4 (2016) 41–57. doi:10.1109/MGRS.2016.2548504.

[15]  J. Buolamwini, Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification *, 2018.

[16]  D. Cousineau, S. Chartier, Outliers detection and treatment: a review., Int. J. Psychol. Res. (2017). doi:10.21500/20112084.844.

[17]  V.J. Hodge, J. Austin, A survey of outlier detection methodologies, Artif. Intell. Rev. (2004). doi:10.1023/B:AIRE.0000045502.10941.a9.

[18]  S. Focardi, F.J. Fabozzi, Black swans and white eagles: On mathematics and finance, Math. Methods Oper. Res. (2009). doi:10.1007/s00186-008-0243-8.

[19]  D.M.J. Tax, One-class classification; Concept-learning in the absence of counter-examples, 2001. doi:10.1063/1.3605545.

[20]  S.S. Khan, M.G. Madden, A survey of recent trends in one class classification, in: Lect. Notes Comput. Sci. (Including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics), 2010. doi:10.1007/978-3-642-17080-5_21.

[21]  S.S. Khan, M.G. Madden, One-class classification: Taxonomy of study and review of techniques, Knowl. Eng. Rev. (2014). doi:10.1017/S026988891300043X.

[22]  A. Patcha, J.M. Park, An overview of anomaly detection techniques: Existing solutions and latest technological trends, Comput. Networks. (2007). doi:10.1016/j.comnet.2007.02.001.

[23]  M.A.F. Pimentel, D.A. Clifton, L. Clifton, L. Tarassenko, A review of novelty detection, Signal Processing. (2014). doi:10.1016/j.sigpro.2013.12.026.

[24]  I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, Y. Bengio, Generative Adversarial Nets (NIPS version), Adv. Neural Inf. Process. Syst. 27. (2014).

[25]  A. Makhzani, Adversarial Autoencoders Alireza, Arxiv. (2016). doi:10.1016/j.msec.2012.07.027.

[26]  M. Mirza, S. Osindero, CGAN, CoRR. (2014).

[27]  D.P. Kingma, V. Kuleshov, Stochastic Gradient Variational Bayes and the Variational Autoencoder, in: ICLR, 2014.

[28]  R. Salakhutdinov, Learning Deep Generative Models, 2015. doi:10.1146/annurev-statistics-010814-020120.