F. Girosi, M. Jones, , and T. Poggio. Priors stabilizers and basis functions: From regularization to radial, tensor and additive splines. Technical Report AIM-1430, Massachusetts Institute of Technology, Cambridge, MA, USA, 1993.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....kernels can be considered as conditionate positive kernels of order 0. Examples of Conditionate positive kernels are : k(x; y) Gamma p kx Gamma yk 2 c 2 Hardy Multiquadric (m = 1) k(x:y) kx Gamma yk 2 ln kx Gamma yk Thin P late Spline (m = 2) As pointed out by Girosi et al. [19] conditionally positive definite kernels are admissible for SVM and in general for all kernel methods. This is underlined by the following result. Theorem 3.3.1.5 Define Q n m the space of polynomials of degree lower than m on R n generates an admissible kernel for SV expansions on the space ....

F.Girosi,M.Jones, and T.Poggio, "Priors, stabilizers and basis functions: From regularization to radial,tensor and additive splines", A.I. Memo No. 1430, MIT, 1993

....n # i (x) # j (x) dx # (20) is smaller than some confidence parameter #. Unfortunately merging of neurons has not been used yet in the simulations described below. 4 Results Gabor and Girosi functions are approximation benchmark. These functions were used previously by Girosi et al. Girosi et al. 1993 ] f gab (x, y) e x 2 cos( 75#(x y) 21) f gir (x, y) sin(2#x) 4(y 0.5) 2 (22) The learning is di#cult because only 20 points were provided for learning from uniformly distributed interval [ 1, 1] 1, 1] for Gabor function (Eq. 21) and from [0, 1] 0, 1] interval ....

....that building the network one should choose the transfer function using only the training error. 1 Methods of preparation of training and testing data and methods of results comparing are the same as used by previous authors. 2 Models 1 to 8 originally published by Girosi, Jones and Poggio in [Girosi et al. 1993] . Model1 P 20 i=1 c i [e (x x i ) 2 # 1 (y y i ) 2 # 2 e (x x i ) 2 # 2 (y y i ) 2 # 1 ] #1 = #2 = 0.5 Model2 P 20 i=1 c i [e (x x i ) 2 # 1 (y y i ) 2 # 2 e (x x i ) 2 # 2 (y y i ) 2 # 1 ] #1 = 10, #2 = 0.5 Model3 P ....

F. Girosi, M. Jones, and T. Poggio. Priors stabilizers and basis functions: From regularization to radial, tensor and additive splines. Technical report, MIT, Cambridge, Massachusetts, 1993.

....For such networks the user would only have to specify the maximum value allowed for the error functional. The network parameters would then be set by this analytic method. The parameters should be optimal. The relation between GRBF networks and spline interpolation is an ongoing research area [13]. Splines are traditionally thought of as weighted polynomials of varying degrees. The weights of these polynomials are set by solving a set of linear equations. A more powerful and general approach is considering splines as a superposition of weighted basis functions. The weights are found by ....

....functional, we should be able to analytically or numerically find the optimal representation, optimal with respect to conditions of continuity and boundedness. Some numerical tests on how different GRBF networks approximate a set of functions are addressed in a paper by Girosi, Jones and Poggio [13]. However, their results are hard to generalize upon. A general and systematic approach to these numerical tests is needed. The tests should be addressed in terms of function spaces, error functionals, and types of discontinuities in the mapping representation (C o , C 1 , and C 2 ) This ....

F. Girosi, M. Jones, and T. Poggio, "Priors, Stabilizers and Basis Functions: from regularization to radial, tensor, and additive splines," A.I. Memo No.1430, Cambridge : Artificial Intelligence Laboratory, Massachusetts Institute of Technology, 1993.

....order and oriented according to the input norm weighting matrix U n . In this sense, the estimate e f constructed above is the smoothest function consistent (up to the regularization parameter ) with the training data. For more details on the deterministic RBF interpolation problem, see [18]. Accepted for publication in IEEE Transactions On Signal Processing Yee and Haykin To compare the two estimator structures, we may rewrite the NWRE general form as e f 0 n (ffl) W n (ffl)Y n where W n (ffl) 4 = W n;j (ffl) n j=1 . Moreover, by substituting (7) into (4) the RBFN ....

M. Jones F. Girosi and T. Poggio, "Priors, stabilizers and basis functions: from regularization to radial, tensor and additive splines", Technical Report A.I. Memo No.1430, MIT, Mar. 1994.

....0.45 0.39 Fuzzy INET 0.18 0.24 Fuzzy VINET 0.076 0.18 stable. 4 To compare some results we will use the root mean square error (RMSE) root of MSE) defined as follows: RMSE = p 1 N P N i=1 (f(x i ) y i ) Gabor and Girosi functions are another approximation benchmark considered here [8]: f gab (x, y) e x 2 cos( 75#(x y) 14) f gir (x, y) sin(2#x) 4(y 0.5) 2 (15) For learning only 20 points were used from uniformly distributed interval [ 1, 1] Theta [ 1, 1] for Gabor function (Eq. 14) and from [0, 1] Theta [0, 1] interval for additive function (Eq. 15) ....

....time reduces sometimes significiently. 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 10 3 10 2 10 1 MSE 0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000 0 1 2 3 4 5 Growth Pruning Figure 4: Hopeful pruning. 5 Models 1 to 8 originally published by Girosi, Jones and Poggio in [8]. The two spiral problem. 6 is quite hard problem for many neural networks, especially when the network should discover not only the spiral, but also discover the architecture of the net. The data consists of two sets (training and testing) with 194 patterns each for two spirals (see figure ....

F. Girosi, M. Jones, and T. Poggio. Priors stabilizers and basis functions: From regularization to radial, tensor and additive splines. Technical report, MIT, Cambridge, Massachusetts, 1993.

....the samples (y i ; x i ) N i=1 , using a radial basis function with Gaussian centers: f(x) n X ff=1 c ff G(kx Gamma t ff k) 1) where G(x) e Gamma x 2 oe 2 (2) and the t 0 ff s are arbitrary parameters termed centers. A radial basis function as defined in Equation 1 was shown in [10] to be a type of regularization network which incorporates a priori knowledge about the smoothness of the function that can be learned from a set of samples. Such a smoothness prior is necessary because the problem of generalizing a function from a set of samples is inherently ill posed, and many ....

F. Girosi, M. Jones, and T. Poggio, "Priors, Stabilizers, and Basis Functions: From Regularization to Radial, Tensor, and Additive Splines", MIT AI Lab Memo, No. 1430, June, 1993.

No context found.

F. Girosi, M. Jones, , and T. Poggio. Priors stabilizers and basis functions: From regularization to radial, tensor and additive splines. Technical Report AIM-1430, Massachusetts Institute of Technology, Cambridge, MA, USA, 1993.

No context found.

Federico Girosi, Michael Jones, and Tomaso Poggio. Priors stabilizers and basis functions: From regularization to radial, tensor and additive splines. Technical Report AIM-1430, 1993.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC