not defined

no text concepts found

CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS GSUA20I6 De Montfort University, Leicester, UK Multi-granularity Bidirectional Cognitive Computing for Uncertain Data Processing 面向不确定性数据处理的多粒度双向认知计算 Guoyin Wang (王国胤) Chongqing Key Lab. of Computational Intelligence, Chongqing University of Posts and Telecommunications, China Inst. of Electronic Information Tech., Chongqing Inst. of Green & Intelligent Tech., CAS, China [email protected] HTTP://CS.CQUPT.EDU.CN/WANGGY 1 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Outline 01 Big Data 02 Cognition and Cognitive Computing 03 Artificial Intelligence with Uncertainty 04 Bidirectional Cognitive Computing (BCC) 05 Conclusions and Prospects 2 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS 01 Big Data Nature: Special Issue on “Big Data”, 2008 Science: Special Issue on “Dealing with Data”, 2011 “Data is a new class of economic asset, like currency and gold.” Source: World Economic Forum 2012 3 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS 1KB=1024B=210B 1MB=210KB=1024KB=220B 1GB=210MB=1024MB=230B 1TB=210GB=1024GB=240B 1PB=210TB=1024TB=250B 1EB=210PB=1024PB=260B Vasant Dhar, Data science and prediction, Communications of the ACM, 2013, 56(12):64-73. 4 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Spaces and Sciences Physical Space Natural Science Data Space Data Science? Social Space Social Science 5 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS 02 Cognition and Cognitive Computing In science, cognition refers to mental processes. These processes include: • Attention • Memory • Producing and understanding language • Solving problems • Making decisions. S. Coren. Sensation & Perception. Harcourt College Publish, 1999. 6 Philosophy of mind ——the nature of mind, mental events, mental Cognitivethe Science functions, mental properties, Cognitive psychology —— mental and their relationship to processes: how people think, consciousness perceive, remember, and learn, etc. the physical body, particularly the brain (mind-body problem). G.J. Feist, E.L. Rosenberg. J. Kim. Problems in the Philosophy of Mind. Psychology: Making Connections. McGraw-Hill Oxford: Humanities/Social Sciences/Languages, 2009. Oxford University Press, 1995. Cognitive linguistics： AI is the science of making intelligent machines. Including :robot, language identification, image recognition, natural language processing and expert system, etc. S.J. Russell, P. Norvig. Artificial Intelligence: A Modern Approach, Prentice Hall, 2009. Cognitive neuroscience —biological Cognitive semantics, Cognitive grammar, Cognitive phonology. W. Croft, D. Alan Cruse. Cognitive Linguistics. Cambridge: Cognitive anthropology is Cambridge University Press, 2004. concerned with what people from different groups know and how that implicit knowledge changes the way people perceive the world around them. R. D‘Andrade. The Development of Cognitive Anthropology, Cambridge: Cambridge University Press, 1995. substrates underlying cognition, with a specific focus on the neural substrates of mental processes. How psychological cognitive functions are produced by the brain. Cognitive Science Hexagon M.S. Gazzaniga. The Cognitive Neurosciences III, G. A. Miller. The cognitive revolution: historical perspective. TRENDS in Cognitive Sciences, Vol. 7 No.3, pp:141-144, 2003. 7 The MIT Press,a2004. CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS 03 Artificial Intelligence with Uncertainty “A new research field in artificial intelligence in the 21st century ----Artificial Intelligence with Uncertainty”. A fundamental problem----The expressing and processing of uncertain concepts D.Y. Li, Y. Du. Artificial Intelligence with Uncertainty(1st ed). London: Chapman and Hall/CRC, 2007. 8 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Key features Randomness Fuzziness Incompleteness Uncertainty Unstableness Inconsistence …… 9 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Unidirectional Computational Cognition Data Unidirectional Transformation Knowledge Common Characteristic Knowledge Discovery Knowledge discovery is the creation of knowledge from structured (relational databases, XML) and unstructured (text, documents, images) sources. Machine Learning Machine learning is a scientific discipline concerned with the design and development of algorithms that allow computers to evolve behaviors based on empirical data, such as from sensor data or databases. Data Mining Data mining (which is the analysis step of Knowledge Discovery in Databases) focuses on the discovery of (previously) unknown properties from data. 10 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Human brain Computer Cognitive Transformation ID Intension of concept Extension of concept Forward Cloud transformation(FCT) C(25, 3, 0.3) …. Birth date ? Backward Cloud Transformation(BCT) 1 0.9 Certainty degree Young people Name 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 Cloud Model 0 10 15 20 25 30 35 40 Cloud drops x 11 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS 04 Bidirectional Cognitive Computing (BCC) Cloud Model Experiments on Bidirectional Cognition Processes 12 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Numerical Characters of Cloud Model Three numerical characters of cloud represent a concept as a whole. Expected value Ex — the expectation of cloud drops in the universe. 0.9 0.8 0.7 0.6 (x) Entropy En — describes the uncertainty measurement of the qualitative concept. 1 He 0.5 0.4 0.3 0.2 0.1 0 10 3En 15 20 Ex 25 30 Hyper entropy He Ex =25, En=3, He=0.3 — the uncertainty measurement of En, and measure whether the concept can be formed. 35 40 x 13 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Cloud Transformation Forward Cloud Transformation (FCT) Forward cloud transformation is a mapping from quality (intension) to quantity (extension). It generates cloud drops according to (Ex, En, He). C (Ex, En, He) 1 0.9 Ex yi=RN(En, He) 0.8 n=1000 0.7 Certainty degree Young people C(25, 3, 0.3) FCT 0.6 0.5 0.4 0.3 0.2 0.1 0 10 15 20 25 30 Cloud drops x (age of young people) 35 40 xi=RN(Ex, yi) (i=1, 2, …, n) x1, x2, …, xn The process of 2nd-order forward cloud transformation (2nd FCT） 14 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS ► Algorithm 2nd order FCT Input: (Ex, En, He), and the number of cloud drops n; Output: n of cloud drops x and their certainty degree , i.e. Drop(xi, i), i = 1,2,…,n; Steps: (1) Generate a normally distributed random number Eni with expectation En and variance He2, i.e. Eni = NORM(En, He2); (2) Generate a normally distributed random number xi with expectation Ex and variance Eni , i.e. xi = NORM(Ex, Eni2); (3) Calculate i e (xi Ex )2 2(En 'i )2 ; (4) xi with certainty degree of i is a cloud drop in the domain; (5) Repeat steps (1) to (4) until n cloud drops are generated. 15 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Backward Cloud Transformation (BCT) Backward cloud transformation is the model for transforming from the quantitative values (extension) to a qualitative concept (intension). It maps a quantity of precise data to a qualitative concept expressed by (Ex, En, He). There are five kinds of backward cloud algorithm as follows: 1 0.9 0.8 Certainty degree 0.7 0.6 BCT 0.5 0.4 0.3 0.2 Young people C(25, 3, 0.3) 0.1 0 10 15 20 25 30 35 40 Cloud drops x (age of young people) 16 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Backward Cloud Transformation (BCT) Backward cloud transformation is the model for transforming from the quantitative values (extension) to a qualitative concept (intension). It maps a quantity of precise data to a qualitative concept expressed by (Ex, En, He). There are five kinds of backward cloud algorithm as follows: ► SBCT-1stM------(C.Y. Liu, D.Y. Li) One-step ► SBCT-4thM------(L.X. Wang) ► MBCT-SD------(G.Y. Wang, C.L. Xu) Multi-steps 2nd order (Ex, En, He) ► MBCT-SR------(G.Y. Wang, C.L. Xu) High order ► pth-BCT(p>2)------(G.Y. Wang, C.L. Xu) (Ex, En1,…, Enp-1, He) 17 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS • Algorithm SBCT-1stM SBCT-1stM is based on the sample variance and the first-order sample absolute central moment to estimate the entropy En and hyper entropy He. 1 n S ( xi X )2 n 1 i 1 2 ② variance Sample Data ① mean （x1, x2,…, xn） ② The 1st-order sample absolute central moment Êx S 2 =En2 +He 2 , ③ E|X -X |= 2 En. ④ ˆ S 2 En 2 He 1 n ˆ En xi Ex , 2 n i 1 1 n E|X -X |= |xi X | n i 1 SBCT-1stM 18 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS ► Algorithm SBCT-1stM Input: Samples x1 , x2 , …., xn . Output: The estimates of a qualitative concept (Ex, En, He). Steps: (1) Calculate the mean, variance and the first-order absolute central moment of sample xi , respectively i.e. 1 n 1 n 1 n 2 2 X xi，S |xi X | . ( xi X ) and E|X -X |= n n i 1 n 1 i 1 i 1 (2) According to the following equations: S 2 =En 2 +He 2 , E|X -X |= 2 En. Calculate the estimates of En and He respectively, i.e: ˆ X , En ˆ Ex 1 n ˆ S 2 En 2 . xi Ex , He 2 n i 1 C.Y. Liu, M. Feng, X.J. Dai, D.Y. Li, “A new algorithm of backward cloud,” Journal of System Simulation, vol. 16, no. 11, pp. 2417-2420, 2004. 19 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS • Characteristics of SBCT-1stM: SBCT-1stM is based on the sample variance and the firstorder sample absolute central moment to estimate the entropy En and hyper entropy He. 1 0.9 0.8 SBCT-1stM Certainty degree 0.7 n=1000 Sample 0.6 0.5 (25.02, 2.96, 0.35) 0.4 0.3 0.2 0.1 0 10 15 20 25 30 35 40 Cloud drops x (age of young people) • Shortage of SBCT-1stM When He is very small, SBCT-1stM may fail to obtain the estimation of He and En. When He 1 > , SBCT-1stM will have a large error for the estimation of He En 3 and En. 20 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS • Algorithm SBCT-4thM SBCT-4thM is based on the sample variance and the fourth-order sample central moment to estimate the entropy En and hyper entropy He. ② variance S2 Sample mean ① mean （x1, x2,…, xn） 1 n ( xi X )2 n 1 i 1 2 Êx ③ 4 = 2 S =En +He , =9He 4 +18He 2 En2 +3En4 4 ② The 4th-order sample central moment 2 ④ ˆ S 2 En 2 He ˆ En 4 9(S 2 ) 2 - 4 6 1 n ( xi X ) 4 n 1 i 1 SBCT-4thM 21 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS ► Algorithm SBCT-4thM Input: Samples x1 , x2 , …., xn . Output: The estimates of a qualitative concept (Ex, En, He). Steps: (1) Calculate the mean, variance and the fourth-order central moment of sample xi , respectively i.e. 1 n 1 n 1 n 2 2 X xi , S ( xi X ) and 4 = ( xi X ) 4 . n i 1 n 1 i 1 n 1 i 1 (2) According to the following equations: S 2 =En2 +He2 , 4 =3(3He4 +6He2 En2 +En4 ). Calculate the estimates of En and He respectively, i.e: ˆ X , En ˆ Ex 4 9(S 2 )2 - 4 6 ˆ S 2 En ˆ 2. , He L.X. Wang. The basic mathematical properties of normal cloud and cloud filter, Personal Communication, May. 2011. 22 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS • Characteristics of SBCT-4thM: SBCT-4thM is based on the sample variance and the fourth-order sample central moment to estimate the entropy En and hyper entropy He. 1 0.9 0.8 Certainty degree 0.7 n=1000 Sample SBCT-4thM 0.6 0.5 (25.02, 2.97, 0.34) 0.4 0.3 0.2 0.1 0 10 15 20 25 30 35 40 Cloud drops x (age of young people) • Shortage of SBCT-4thM In Step 2, SBCT-4thM may fail to estimate the entropy En and 9(S 2 )2 - 2 ˆ 2 <0, for example, n hyper entropy He when <0 and S En 4 6 is a small number. 23 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS ► Algorithm MBCT-SD Input: Samples x1 , x2 , …., xn , group number m, each group sample size r. Output: The estimates of a qualitative concept (Ex, En, He). Steps: n 1 ˆ x . (1) Calculate the sample mean of sample xi , i.e. Ex i n i 1 (2) Obtain the new sample from sample xi , that is, drawing m groups sample from xi randomly and each group has r samples, and n=m*r, n, m, r are positive integers. Calculate each group sample variance, i.e. 1 r 1 r 2 ˆ ˆ yˆ ( xij Exi ) , where, Exi xij (i 1, 2, r 1 j 1 r j 1 2 i , m) 2 2 (3) Calculate the estimates of En2 and He2 from the new sample y1 , y2 , respectively, i.e: , y m2 ˆ 2 1 4( EY ˆ 2 )2 2 DY ˆ 2 , He ˆ 2 EY ˆ 2 En ˆ 2. En 2 m 1 1 m 2 ˆ 2 2 2 2 ˆ 2 ˆ where, EY yˆi , DY ( yˆi EY ) . m i 1 m 1 i 1 G. Y. Wang, C.L. Xu, Q.H. Zhang, X.R. Wang. A Multi-step Backward Cloud Generator Algorithm, RSCTC2012, LNAI 7413, pp: 313-322, 2012. 24 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS • Algorithm MBCT-SR C ( Ex, En, He) FCT x1, x2,……, xn xx1r ,; 11 12 n 1 Sample with ˆ x . Ex i replacement n i 1 xi=RN(Ex, yi) yi=RN(En, He) …, ˆ 2 , He ˆ 2 En y12, y22,…, ym2 y12 x21, x22,…, x2r; …; xm1, xm2,…, xmr . y2 2 … ym2 • MBCT-SR is also a two-step method to estimate the entropy En and hyper entropy He. 25 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS ► Algorithm MBCT-SR Input: Samples x1 , x2 , …., xn , group number m, each group sample size r. Output: The estimates of a qualitative concept (Ex, En, He). Steps: 1 n X xi . (1) Calculate the sample mean of sample xi , i.e. n i 1 (2) Obtain the new sample from sample xi , that is, drawing m groups sample with replacement from xi randomly and each group has r samples (n, m, r are positive integers). Calculate each group sample variance, i.e. r 1 r 1 2 ˆ ) , where, Ex ˆ x (i 1, 2, yˆ ( xij Ex i i ij r 1 j 1 r j 1 2 i , m) (3) Calculate the estimates of En2 and He2 from the new sample respectively, i.e: y12 , y22 , , y m2 ˆ 2 1 4( EY ˆ 2 )2 2 DY ˆ 2 , He ˆ 2 EY ˆ 2 En ˆ 2. En 2 1 m 2 ˆ 2 1 m 2 ˆ 2 2 2 ˆ where, EY yˆi , DY ( yˆi EY ) . m i 1 m 1 i 1 Chang Lin XU, Guo Yin WANG, Backward Cloud Transformation Algorithm for Realizing Stability Bidirectional Cognitive Mapping, PR&AI, 2013, 26(7):634-642. 26 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS ► Algorithm pth order BCT (p>2) —— pth order Backward Cloud Transformation Algorithm Input: Samples x1 , x2 , …., xn , group number m, each group sample size r. Output: The estimates of (Ex=En1, En2, En3, ……, Enp-1, Enp, He). Steps: n 1 (1) Calculate the sample mean of sample xi , i.e. X xi . n i 1 (2) Obtain the new sample from sample xi , that is, drawing m groups sample with replacement from xi randomly and each group has r samples (n, m, r are positive integers). Calculate each group sample variance, i.e. 1 r 1 r 2 ˆ ˆ yˆ ( xij Exi ) , where, Exi xij (i 1, 2, r 1 j 1 r j 1 2 i Let Y y12 , y22 , , m) , y m2 . Guoyin Wang, Changlin Xu,Qinghua Zhang,Xiaorong Wang, p-order Normal Cloud Model Recursive Definition and Analysis of Bidirectional Cognitive Computing, Chinese Journal of Computers, 2013, 36(11):2316-2329. 27 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS (3) Calculate the estimates of En2 , sample respectively, i.e: , En2p1 , En 2p , He2 from the new 2 for ( p =2; p<P; p++) ˆ 2 1 4( EY ˆ 2 )2 2 DY ˆ 2, { En p 2 1 m 2 ˆ 2 1 m 2 ˆ 2 2 2 ˆ where, EY yˆi , DY ( yˆi EY ) m i 1 m 1 i 1 Input a group new value: mm, rr, and let m=mm, r=rr; ˆ 2 1 4( EY ˆ 2 )2 2 DY ˆ 2 , En p 1, i p 1, i p 1, i 2 ˆ 2 En ˆ 2 , (i 1, 2, , m) uˆ 2p 1,i EY p 1, i p 1, i r r ˆ 2 1 yˆ 2 , DY ˆ 2 1 ( yˆ 2 EY ˆ 2 )2 where, EY p 1, i j p 1,i j p 1,i r j 1 r 1 j 1 Let ,Y uˆ 2p 1,1 , uˆ 2p 1,2 ,......, uˆ 2p 1, m ; } Estimate the En2p , He2p from the sample set Y, i.e. ˆ 2 1 4( EY ˆ 2 )2 2 DY ˆ 2 , He ˆ 2 EY ˆ 2 En ˆ 2, En p p p 2 1 m 2 ˆ 2 1 m 2 ˆ 2 2 2 ˆ where, EY yˆi , DY ( yˆi EY ) . m i 1 m 1 i 1 28 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Generic Normal Cloud Transformation 2nd-order Generic Forward Cloud Transformation (2nd order G-FCT) 2nd-order Generic Backward Cloud Transformation (2nd order G-BCT) pth-order Generic Forward and Backward Cloud Transformation (pth order G-FCT, pth order G-BCT) 29 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS 2nd order G-FCT 2nd order FCT Concept (i=1, 2, …, n) C (Ex, En, He) Ex yi=RN(En, He) 2nd order G-FCT Concept C (Ex, En, He) Ex yi=RN(En, He) (i=1, 2, …, m) xij=RN(Ex, yi) xi=RN(Ex, yi) Cloud Drops x1, x2, …, xn The process of 2nd order FCT (j=1, 2, …, ri) Cloud Drops x11, x12, …, x1r1; x21, x22, …, x2r2; …… ; xm1, xm2,…, xmrm; The process of 2nd order G-FCT 30 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS ► Algorithm 2nd order G-FCT Input: (Ex, En, He), and the number of cloud drops n; the parameters m, ri (i=1, 2 ,… , m) Output: n cloud drops xij with certainty degree (xij ) (i=1,2,…,m; j=1,2,…,ri) Steps: (1) Generate m normally distributed random numbers yi with expectation En and variance He2, i.e. yi = NORM(En, He2); (2) For each y in step 1, generate ri normally distributed random numbers xij with expectation Ex and variance yi , i.e. xi = NORM(Ex, yi2); (3) Calculate ( xij ) e (xij Ex )2 2yi2 ; (4) xij with certainty degree (xij ) is a 2nd-order generic normal cloud drop in the domain; (5) Repeat steps (1) to (4) until n cloud drops are generated. G.Y. Wang, C.L. Xu, D.Y. Li . Generic Normal Cloud Model, Information Sciences, (2014) 31 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Let (Ex, En, He)=(25, 3.0, 0.55), r1= r2= … = rm=r, n=5000 . When m, r take different values respectively (m*r=n), the shapes of the 2nd order G-FCT are as follows 0.6 1 0.8 0.8 0.6 0.6 0.6 1 0.8 1 2nd order 0.8 FCT 1 0.4 0.4 0.4 0.4 0.2 0.2 0.2 0.2 0 10 20 30 0 10 40 20 30 0 10 40 20 x x (a) m=5000, r=1 30 0 10 40 20 x (b) m=500, r=10 30 40 x Normal (d) m=50, r=100 (c) m=100, r=50 Distribution 1 1 0.8 0.8 0.8 0.8 0.6 0.6 0.6 0.6 1 1 0.4 0.4 0.4 0.4 0.2 0.2 0.2 0.2 0 10 20 30 x (e) m=10, r=500 40 0 10 20 30 x (f) m=5, r=1000 40 0 10 20 30 x (g) m=2, r=2500 40 0 10 30 32 20 40 x (h) m=1, r=5000 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Conclusions ① When m=n, r1= r2= … = rm =1, for each yi, only one cloud drop xij will be obtained, that is, a cloud drop xij corresponds to a yi, then n cloud drops are from n different normal distributions respectively, so, the 2nd order G-FCT will be a 2nd order FCT proposed in reference [1] ② When m=1, r1=n, there will be only one y1=RN(En, He), then the 2nd order G-FCT will be a normal distribution N(Ex, y1). [1] D.Y. Li, Y. Du. Artificial Intelligence with Uncertainty(1st ed). London: Chapman and Hall/CRC, 2007. 33 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Experiments on Bidirectional Cognition Processes (1) Cognizing a concept over and over again (2) Cognizing process with the increasing of sample size n (3) Many people’s mutual cognizing process for a concept (4) Multi granularity concept cognition (5) Image segmentation Based on Bidirectional Computational Cognition 34 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS (1) Cognizing a concept over and over again 1 FCT 0.9 0.8 0.7 BCTi C(Exi, Eni, Hei) C(Ex, En, He) 0.6 L times 0.5 0.4 0.3 0.2 0.1 0 -15 BCTi ☺ -10 -5 0 x 5 10 Simulate the cognitions of SBCT-1stM, SBCT-4thM, MBCT-SD, MBCT-SR and 3rd-BCT respectively with different cycle number L . 35 15 1 0.9 C1(24.96, 2.98, 0.28) When the number of cycles L=1 Certainty degree SBCT-1stM 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 15 17 19 21 23 25 27 29 31 33 35 Cloud drops x (Ages of young people) 1 0.9 SBCT-4thM C2(25.02, 2.95, 0.21) Certainty degree 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 15 0.7 0.6 MBCT-SD 0.5 C3(24.98, 3.01, 0.09) L =1 0.4 0.3 Certainty degree Certainty degree 21 23 25 27 29 31 33 35 (Ages of young people) 0.9 0.8 0.2 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.1 0 15 17 35 Conclusion: When the qualitative concept is quite clear, the cognition results of SBCT1stM, SBCT-4thM and 3rd-BCT have some excursion，while the MBCT-SD and MBCT-SR’s cognition results are very good. MBCT-SR C4(25.03, 2.97, 0.09) 3rd-BCT C5(25.02, 2.98, 0.32, 0.02) Certainty degree Clear concept 19 21 23 25 27 29 31 33 35 Cloud drops x (Ages of young people) Certainty degree 0 15 17 19 21 23 25 27 29 31 33 Cloud drops x (Ages of young people) 19 1 0.9 C(25, 3, 0.1) 17 Cloud drops x 1 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 15 17 19 21 23 25 27 29 31 33 35 Cloud drops x 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 15 17 19 21 23 25 27 29 31 33 35 Cloud drops x 36 1 0.9 SBCT-1stM C1(24.92, 2.96, 0.34) When the number of cycles L=50 Certainty degree 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 16 18 20 22 Cloud drops x 24 26 28 30 32 34 (Ages of young people) 1 0.9 Certainty degree 0.8 SBCT-4thM 0.7 0.6 0.5 0.4 0.3 0.2 C2(25.03, 2.98, 0.30) 0.1 0 16 18 20 22 1 1 0.9 0.9 0.8 0.8 0.7 0.6 0.5 L =50 0.4 MBCT-SD C3(24.98, 3.01, 0.09) 0.3 Certainty degree C(25, 3, 0.1) Certainty degree Cloud drops x 26 28 30 32 34 36 (Ages of young people) 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0.2 0 15 0.1 MBCT-SR C4(24.97, 2.85, 0.091) Conclusion: When the qualitative concept is quite clear, the cognition results of SBCT-1stM, SBCT-4thM and 3rd-BCT have some excursion，while the MBCT-SD and MBCT-SR’s cognition results are very good. 3rd-BCT C5(24.93, 2.97, 0.29, 0.009) Certainty degree Clear concept 17 19 21 23 25 27 29 31 33 35 Cloud drops x (Ages of young people) 35 Certainty degree 0 15 17 19 21 23 25 27 29 31 33 Cloud drops x (Ages of young people) 24 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 15 17 19 21 23 25 27 29 31 33 35 Cloud drops x 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 37 0 15 17 19 21 23 25 27 29 31 33 35 Cloud drops x CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS • Conclusions ▬ MBCT-SD and MBCT-SR over perform SBCT-1stM, SBCT-4thM and 3rd-BCT for cognizing uncertain concepts in all cases. ▬ MBCT-SD and MBCT-SR could be used to construct stable bidirectional cognitive mapping between concept’s intension and extension together with FCT. ▬ The 5 Backward Cloud Transformations could be used to cognize a concept from extension to intension. They can estimate different kinds of people respectively. 38 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Experiments on Bidirectional Cognition Processes (1) Cognizing a concept over and over again (2) Cognizing process with the increasing of sample size n (3) Many people’s mutual cognition process for a concept (4) Multi granularity concept cognition (5) Image segmentation Based on Bidirectional Computational Cognition 39 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS ？ ？ C(Ex, En, He) When the sample size n is increasing constantly ？ ？ ？ 40 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS 1 0.9 Clear concept Certainty degree 0.8 C(25, 3, 0.1) 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 15 17 19 21 23 25 27 29 31 33 35 Cloud drops x (Ages of young people) 1 0.9 0.8 Certainty degree Uncertain concept C(25, 3, 0.55) 0.7 0.6 0.5 0.4 0.3 0.2 0.1 Confusing concept C(25, 1, 0.8) Certainty degree 0 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 Cloud drops x 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 15 17 19 21 23 25 27 29 31 33 35 Cloud drops x 41 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS 1 0.9 0.8 Certainty degree • The cognizing process of MBCT-SR when the sample size n is increasing C(25, 3, 0.55) 0.7 0.6 0.5 0.4 0.3 0.2 Uncertain concept C4(20, 0, 0) 0.4 0.2 12 14 16 18 20 22 24 26 28 30 Cloud drops x C4(24.74, 3.15, 0.78) 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 Cloud drops x C4(24.91, 3.03, 0.53) Certainty degree Certainty degree n=50 1 n=400 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 Cloud drops x 1 n=10 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 Cloud drops x C4(25.02, 3.01,0.54) Certainty degree 0.6 1 n=2 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 Cloud drops x C4(24.37, 2.14, 0.80) Certainty degree Certainty degree Certainty degee n=1 0.8 0 10 0 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 Cloud drops x C4(25.87, 1.92, 1.21 ) 1.2 1 0.1 1 n=1000 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 Cloud drops x 42 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS • Conclusions ▬ The cognition processes of the 5 backward cloud transformations have some differences when the sample size n is increasing. For a clear concept, MBCT-SD and MBCT-SR have good cognition results, while the cognition results of SBCT-1stM, SBCT-4thM and 3rd-BCT have some excursion. For a confusing concept, SBCT-1stM, SBCT-4thM, MBCT-SD and MBCT-SR have good cognition results, while the cognition result of 3rd-BCT has much excursion. ▬ SBCT-1stM, SBCT-4thM and MBCT-SD may fail to cognize a concept when there are only few samples (a very small n). 43 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Experiments on Bidirectional Cognition Processes (1) Cognizing a concept over and over again (2) Cognizing process with the increasing of sample size n (3) Many people’s mutual cognition process for a concept (4) Multi granularity concept cognition (5) Image segmentation Based on Bidirectional Computational Cognition 44 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS (3) Many people’s mutual cognition process for a concept Simulate the cognizing process when a qualitative concept is passed from one person to another over and over again. ..... 45 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Experiment method： Ci(Exi, Eni, Hei) FCT ↔BCTi FCT ↔BCTj C(Ex, En, He) Cj(Exj, Enj, Hej) …… FCT ↔BCTk Ck(Exk, Enk, Hek) 46 1 0.9 Certainty degree 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 Cloud drops x MBCT-SD C(25, 3, 0.55) 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 Cloud drops x SBCT-4thM C3(25.02, 3.02, 0.54) C2(24.97, 2.97, 0.52) SBCT-1stM MBCT-SR C4(25.03, 2.98, 0.56) C1(24.98, 3.04, 0.53) 1 0.9 0.8 Certainty degree Uncertain concept 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 Cloud drops x Certainty degree Certainty degree 0.8 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 Cloud drops x Certainty degree CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 Cloud drops x 47 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS • Conclusions ▬ A concept could be passed among different kinds of people(BCTs) with excursion in some degree. 48 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Experiments on Bidirectional Cognition Processes (1) Cognizing a concept over and over again (2) Cognizing process with the increasing of sample size n (3) Many people’s mutual cognition process for a concept (4) Multi granularity concept cognition (5) Image segmentation Based on Bidirectional Computational Cognition 49 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Prof. Leslie Valiant at Harvard University Turing Award 2010 Current Research Interests “A fundamental question for artificial intelligence is to characterize the computational building blocks that are necessary for cognition.” Information/Knowledge Granules http://people.seas.harvard.edu/~valiant/researchinterests.htm 50 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Members age clustering of Chinese Academy of Engineering 0.07 0.06 0.05 C2(76.4,5.9,0.29) 0.04 0.03 0.02 C1(53,3.3,0.16) 0.01 0 40 50 60 70 age 80 90 100 Two concepts generated by A-GCT Concept tree generated by A-GCT at different granularity level 51 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Deep-Learning • Based on connectionism in the 1980's. • Adds the assumption: factors are organized into multiple levels of abstraction or composition • Deep architectures are composed of multiple levels of non-linear operations G. E. Hinton, R. R. Salakhutdinov, Reducing the Dimensionality of Data with Neural Networks, Science, 313:504-507, 2006 G.E. Hinton, S. Osindero, Y. Teh, A fast learning algorithm for deep belief nets, Neural Computation, 18:1527-1554, 2006 Y. Bengio, P. Lamblin, D. Popovici, H. Larochelle, Greedy Layer-Wise Training of Deep Networks, NIPS 2006, pp. 153-160 M.A. Ranzato, etc., Efficient Learning of Sparse Representations with an Energy-Based Model, NIPS 2006, pp. 1137-1144 Y. Bengio, Learning Deep Architectures for AI, Foundations and Trends in Machine Learning, 2(1):1-127, 2009 G. Anthes, Deep Learning Comes of Age, Communications of the ACM, 56(6): 13-15,2013 52 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS ANFIS : adaptive-network-based fuzzy inference system A1 B1 f1 p1 x q1 y r1 f A2 x Y X w1 f1 w2 f 2 B2 X w1 f1 w2 f 2 w1 w2 Y y Fuzzy Inference (Type 3) J.-S. R. Jang, “ANFIS: Adaptive-network-based fuzzy inference system,” IEEE Trans. Syst., Man, Cybern., vol. 23, pp. 665–685, 1993. 53 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS w1 f1 w2 f 2 f w1 w2 w1 f1 w2 f 2 x y A1 x A2 П W1 N W1 W1 f1 Σ f B1 П y B2 W2 N W2 f2 W2 x y ANFIS 54 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS – TMLNN: Triple-Valued or Multiple-Valued Logic Neural Network X1 Classical neuron X2 W1 W2 …… Σ θ Y=f(Σ(WiXi)-θ) Wn Xn X1 Triple-valued or multiple-valued logic neuron (TMLN) X2 W1 W2 …… I f(I) Y=g(f(I)) Wn Xn G.Y. Wang, H.B. Shi, Three Valued Logic Neural Network, Proc. of Int. Conf. on Neural Information Processing, Hong Kong, 1112-1115, 1996. G.Y. Wang, H.B. Shi, TMLNN: Triple-Valued or Multiple-Valued Logic Neural Network, IEEE Trans. on Neural Networks, 9(6):1099-1117, 1998. 55 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Multiple-Valued Logic “Exclusive OR(XOR)” Experiment A B XOR (A, B) XOR -1 -1 -1 -4.5 -1 0 0 -1 1 1 0 -1 0 0 0 0 0 1 0 1 -1 1 1 0 0 1 1 -1 1 N_XOR 2.5 2.6 N_NAB A 0 N_ANB -1.9 ANB NAB 1.6 A -1.5 1.5 -1 -1.8 -1 B 0 1 -1.6 B 56 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS MGrC for Big Data Traditional DM: Granularity Space Optimization ----Searching a suitable granularity level for problem solving. Traditional GrC: Granularity Level Switching ----Solving a problem in different granularity levels. New Direction: Multi-Granularity Joint Problem Solving ----Solving a problem in Multi-Granularity levels jointly and simultaneously. 57 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Experiments on Bidirectional Cognition Processes (1) Cognizing a concept over and over again (2) Cognizing process with the increasing of sample size n (3) Many people’s mutual cognition process for a concept (4) Multi granularity concept cognition (5) Image segmentation Based on Bidirectional Computational Cognition 58 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Image segmentation The original images 59 Results of image segmentation Fig.1 Expected segmentation results Fig.2 Segmentation results by K-means Fig.3 Segmentation results by SBCT-1stM* Fig.4 Segmentation results by MBCT-SR * Kun Qin, Kai Xu, Yi Du, Deyi Li. An Image Segmentation Approach Based on Histogram Analysis Utilizing Cloud Model. 2010 Seventh International Conference on Fuzzy Systems and Knowledge Discovery (FSKD 2010):524-528. 60 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Results of image segmentation Error Rate (%) Method Bird and branch Starfish Flower Butterfly Leopard and branch K-means 69.335 52.183 29.761 45.122 47.558 SBCT1stM 29.756 32.946 16.423 36.901 44.246 MBCT-SR 10.401 10.939 11.462 15.154 19.126 61 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Test on noisy images The images with noise (5% salt & pepper noise) 62 The test of noise immunity Fig.1 Expected segmentation results Fig.2 Segmentation results by K-means Fig.3 Segmentation results by SBCT-1stM Fig.4 Segmentation results by MBCT-SR 63 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Test on noisy images Error Rate(%) Method Bird and branch Starfish Flower Butterfly Leopard and branch K-means 86.535 74.132 68.743 74.571 68.240 SBCT-1stM 60.041 46.165 16.094 85.997 43.563 MBCT-SR 11.968 12.791 12.603 21.861 40.552 64 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS 0.015 frequence 0.01 0.005 0 0 50 100 150 gray 200 250 300 Gray histogram laser cladding image 0.015 frequence 0.01 0.005 0 0 50 100 150 gray 200 250 300 Three concept generated by A-GCT: C1(81.3, 15.3, 1.81), C2(172.3, 31.7,3.74), C3(253.2, 1.3, 0.11) 65 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS Medical Image Segmentation [1] C.M. Li. Distance Regularized Level Set Evolution and Its Application to Image Segmentation, IEEE Trans. Image Process. 19 (12) (2010) 3243-3254. [2] Z. Zivkovic. Gentle ICM energy minimization for Markov random fields with smoothness-based priors, Journal of Real-Time Image Processing. 11 (1) (2016) 235-246. [3] S.F. Dai, et al. A novel approach of lung segmentation on chest CT images using graph cuts, Neurocomputing. 1 (2015) 799-807. [4] M.B. Salah, A. Mitiche, I.B. Ayed, Multiregion Image Segmentation by Parametric Kernel Graph Cuts, IEEE Trans. Image Process. 20 (2) (2011) 545-557. 66 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS 05 Conclusions and Prospects Conclusions: The relationship of cognitive science and artificial intelligence is studied. Bidirectional cognitive computing (BCC) is proposed. Cloud Model is studied as a case study of BCC. Some human cognition processes are implemented successfully with BCC. BCC is used to solve some real life key problems like image segmentation. 67 CHONGQING UNIVERSITY OF POSTS AND TELECOMMUNICATIONS 68