1. THE NEURAL-NETWORK ANALYSIS & its applications DATA FILTERS Saint-Petersburg State University JASS 2006

2. About me Name: Name: Alexey Minin Place of studying: Place of studying: Saint-Petersburg State University Current semester: Current semester: 7 th semester Field of interests: Field of interests: Neural Nets, Data filters for Optics (Holography), Computational Physics,EconoPhisics.

3. Content: What is Neural Net & it’s applications Neural Net analysis Self organizing Kohonen maps Data filters Obtained results

4. What is NeuroNet & it’s applications Recognition of images Recognition of images Processing of noised signals Processing of noised signals Addition of images Addition of images Associative search Classification Drawing up of schedules Optimization The forecast Diagnostics Prediction of risks

5. What is Neural Net & it’s applications M-X2 9980 Recognition of images

6. What is Neural Net & it’s applications

7. PARADIGMS of neurocomputing Neural Net analysis Connection Localness and parallelism of calculations The training based on data (programming) Universality of training algorithms

8. Neural Net analysis What is Neuron? Typical formal neuron makes the elementary operation – weighs values of the inputs with the locally stored weights and makes above their sum nonlinear transformation: neuron makes nonlinear operation above a linear combination of inputs

9. Neural Net analysis Global communications Formal neurons Layers Connectionism

10. Neural Net analysis Localness and parallelism of calculations Localness of processing of the information Any neuron reacts only to the information from connected with it neurons without the appeal to a general plan of calculations Neurons are capable to function in parallel Parallelism of calculations

11. Comparison of ANN&BNN BRAINPC IBM Vprop=100m/sVprop=3*10 8 m/s 100hz10 9 hz N=10 10 -10 11 neurons N=10 9 The parallelism degree ~10 14 like 10 14 processors with 100 Hz frequency. 10 4 connected at the same time.

12. The training based on data (programming) Neural Net analysis Absence of the global plan Mode of distribution of the information on a network with corresponding adaptation neurons The algorithm is not set in advance, and generated by data Training of a network occurs on a small share of all possible situations then the trained network is capable to function in much wider range of patterns Local change by any neuron the selected parameters Synaptic weights Training of a network Patterns, on which Network is training An ability for generalization

13. Neural Net analysis Universality of training algorithms The only principle of studying - - is to find minimum of empirical error W – set of synaptic weights E (W) – error function The task is to find Global minimum The stochastic optimization as a way not to stick at local minimum

14. Neural Net analysis BASIS NEURAL NETS Perceptron Hopfield network Kohonen maps Probabilistic NNets NN with general regression Polynomial nets

15. The architecture of NN Neural Net analysis LEVEL-BY-LEVEL WITHOUT FEEDBACK RECURRENT with FEEDBACK (Elman-Jordan) PROTOTYPES OF ANY NEURAL ARCHITECTURE

16. Classification of NN Neural Net analysis By type of training with tutorwithout tutor In this case the network is offered most to find the latent laws in data file. So, redundancy of data supposes compression of the information, and a network it is possible to learn to find the most compact representation of such data, i.e. to make optimum coding the given kind of the entrance information.

17. Methodology of self-organizing cards Self-organizing Kohonen cards represent the type of the neural networks trained without the teacher. The network independently forms the outputs, adapting to signals acting on its input. As "teacher" of a network only data, that is an information available in them, the laws distinguishing entrance data from casual noise can serve. Cards unite in themselves two types of compression of the information: Downturn of dimension of data with the minimal loss of the information Reduction of a variety of data due to allocation of a final set of prototypes, and references of data to one of such types

18. Schematic representation of self-organizing network Methodology of self-organizing cards Neurons in the target layer are ordered and correspond to cells of a bi-dimensional card which can be painted by a principle of affinity of attributes

19. Hebb training rule Hebb, 1949 Change of weight at presentation of i th example is proportionally its inputs and outputs: : Change of weight at presentation of i th example is proportionally its inputs and outputs: : If to formulate training as a problem of optimization trained on Hebb neuron aspires to increase amplitude of the output: Where averaging is spent on training sample Training on Hebb in that kind in what it is described above, In practice not useful since leads to unlimited increase of amplitude of weights. NB: in this case there is no minimum error Vector representation

20. Oya training rule The member interfering is added To unlimited growth of weights Vector representation Rule Oya maximizes sensitivity of an output neuron at the limited amplitude of weights. It is easy to be convinced of it, having equated average change of weights to zero. Having increased then the right part of equality on w. We are convinced, that in balance Thus, weights trained neuron are located on hyper sphere : At training on Oya, a vector of weights neuron settles down on hyper sphere, In a direction maximizing Projection of entrance vectors.

21. Competition of neurons: the winner takes away all Basis algorithm Training of a competitive layer remains constant Winner: # of neuron winner I.e. the winner will appear neuron, giving the greatest response to the given entrance stimulus Training of the winner:

22. The winner takes away all One of variants of updating of a base rule of training of a competitive layer Consists in training not only the neuron-winner, but also its "neighbors", though and with In the smaller speed. Such approach - "pulling up" of the nearest to the winner neuron- It is applied in topographical Kohonen cards Function of the neighborhood is equal to unit for the neuron- -winner with an index And gradually falls down at removal from the neuron-winner Modified by Kohonen training rule Training on Kohonen reminds stretching an elastic grid of prototypes on Data file from training sample

23. Bidimentional topographical card of a set Three- dimensional data Each point in three-dimensional space gets in the cell of a grid having coordinate of the nearest to it’s neuron from bidimentional card.

24. The convenient tool of visualization Data is coloring topographical Cards, it is similar to how it do on Usual geographical cards. All attribute of data generates the coloring Cells of a card - on size of average value This attribute at the data who have got in given Cell. Visualization a topographical card, Induced by i-th component of entrance data Having collected together cards of all interesting Us of attributes, we shall receive topographical The atlas, giving integrated representation About structure of multivariate data.

25. Classified SOM for NASDAQ100 index for the period from 10-Nov-1997 till 27-Aug-2001 Methodology of self-organizing cards

26. Change in time of the log- price of actions of companies JP Morgan Chase (The top schedule) and American Express (the bottom schedule) for the period With 10-Jan-1994 on 27-Oct-1997 Change in time of the log- price of actions of companies JP Morgan Chase (The top schedule) and Citigroup (the bottom schedule) for the period c 10-Nov-1997 on 27-Aug- 2001

27. How to choose a variant? This is the forecast of the Sea level (Caspian)

28. DATA FILTERS Custom filters (e.g. Fourier filter) Adaptive filters (e.g. Kalman filter) Empirical mode decomposition Holder exponent

29. Adaptive filters won’t change phase Further we will keep in mind, that we are going to make forecasts, that’s why we need filters, which won’t change phase of the signal. Z -1 X(n) X(n-1) X(n-2) … X(n-nb) Z -1 b(2) b(3) b(nb+1) Z -1 -a(2) -a(3) -a(na+1) y(n) y(n-1) y(n-2) … y(n-nb)

30. Adaptive filters We saved all maxima, there is no phase distortion Siemens value, ad close (scaled)

31. Adaptive filters Let’s try to predict next value using zero-phase filter, having information about historical price: I used Perceptron with 3 hidden layers, logistic act function, rotation alg, 20 min

32. Adaptive filters Kalman filter K(n) ++ Z -1 ac

33. Adaptive filters Lets use Kalman filter, like the error estimator for the forecast of the zero-phase filtered data.

34. Empirical Mode Decomposition What is it? We can heuristically define a (local) high-frequency part {d(t), t− ≤ t ≤ t+}, or local detail, which corresponds to the oscillation terminating at the two minima and passing through the maximum which necessarily exists in between them. For the picture to be complete, one still has to identify the corresponding (local) low-frequency part m(t), or local trend, so that we have x(t) = m(t) + d(t) for t− ≤ t ≤ t+.

35. What is it? Empirical Mode Decomposition

36. Algorithm Given a signal x(t), the effective algorithm of EMD can be summarized as follows: 1. identify all extrema of x(t) 2. interpolate between minima (resp. maxima), ending up with some envelope emin(t) (resp. emax(t)) 3. compute the mean m(t) = (emin(t)+emax(t))/2 4. extract the detail d(t) = x(t) − m(t) 5. iterate on the residual m(t)

72. Empirical Mode Decomposition Lets do it for Siemens index

73. Empirical Mode Decomposition Lets do it for Siemens index We saved all strong maxima and there is no phase distortion

74. Empirical Mode Decomposition Lets make a forecast for Siemens index THERE WAS NO DELAY IN THE FORECAST AT ALL!!!

75. Holder exponent The main idea is next. Consider Holder derived, that So this formula is a somewhat connection between “bad” functions and “good” functions. If we will look on this formula with more precise we will notice, that we can catch moments in time, when our function knows, that it’s going to change it’s behavior from one to another. It means that today we can make a forecast on tomorrow behavior. But one should mention that we don’t know the sigh on what behavior is going to change.

76. Results

77. Thank You! Any QUESTIONS? SUGGESTIONS? IDESAS? Soft I’m using: 1)MatLab 2)NeuroShell 3)FracLab 4)Statistika 5)Builder C++