Radial-Basis Function etworks A function is radial basis () if its output depends on (is a non-increasing function of) the distance of the input from a given stored vector. s represent local receptors, as illustrated below, where each point is a stored vector used in one. In a network one hidden layer uses neurons with activation functions describing local receptors. Then one output node is used to combine linearly the outputs of the hidden neurons. w3 P w w The vector P is interpolated using the three vectors; each vector gives a contribution that depends on its weight and on its distance from the point P. In the picture we have w < w3 < w 5 ARCHITECTURE x ϕ w x y ϕ m w m x m One hidden layer with activation functions ϕ... ϕ m Output layer with linear activation function. y w ϕ ( x t ) +... + w m ϕ m ( x t m x t distance of x ( x,..., xm ) from vector t ) 5 Rossella Cancelliere
HIDDE EURO MODEL Hidden units: use radial basis functions φ σ ( x - t ) the output depends on the distance of the input x from the center t x φ σ ( x - t ) x x m ϕ σ t is called center σ is called spread center and spread are parameters 5 3 Hidden eurons A hidden neuron is more sensitive to data points near its center. For Gaussian this sensitivity may be tuned by adjusting the spread σ, where a larger spread implies less sensitivity. 5 4 Rossella Cancelliere
Gaussian φ φ : center σ is a measure of how spread the curve is: Large σ Small σ 5 5 Interpolation with The interpolation problem: m Given a set of different points { x R, i L } i and a set of real numbers { R, i L }, find a function F : R m R d i that satisfies the interpolation condition: F ( x i ) d i If F ( x ) w iϕ ( x x i ) i ϕ( x x )...... ϕ( x x )... ϕ( x ϕ( x x x we have: ) ) w... w d... d Φ w 5 6 d Rossella Cancelliere 3
Types of φ Micchelli s theorem: Let { x be a set of distinct points in. Then the -by- interpolation i} i R matrix Φ, whose ji-th element is ϕ ji ϕ ( x j x i ) is nonsingular. Multiquadrics: ϕ ( r ) ( r + c ) Inverse multiquadrics: ϕ ( r ) ( r + c Gaussian functions (most used): r ϕ ( r ) exp σ m ) σ > 0 c > 0 r x t 5 7 Example: the XOR problem Input space: x (0,) (,) (0,0) (,0) x Output space: 0 y Construct an pattern classifier such that: (0,0) and (,) are mapped to 0, class C (,0) and (0,) are mapped to, class C 5 8 Rossella Cancelliere 4
Example: the XOR problem In the feature (hidden layer) space: ϕ ( x t ϕ ( x t ) e ) e x t x t with t (,) and t (0,0) φ.0 (0,0) Decision boundary 0.5 (,) (0,) and (,0) 0.5.0 φ When mapped into the feature space < ϕ, ϕ > (hidden layer), C and C become linearly separable. So a linear classifier with ϕ (x) and ϕ (x) as inputs can be used to solve the XOR problem. 5 9 for the XOR problem x t ϕ ( x t ) e with t (,) and t (0,0) ϕ ( x t ) e x t y e x t x t e x t x t - y - + + If y > 0 then classotherwise class 0 5 0 Rossella Cancelliere 5
network parameters What do we have to learn for a with a given architecture? The centers of the activation functions the spreads of the Gaussian activation functions the weights from the hidden to the output layer Different learning algorithms may be used for learning the network parameters. We describe three possible methods for learning centers, spreads and weights. 5 Learning Algorithm Centers: are selected at random centers are chosen randomly from the training set Spreads: are chosen by normalization: Maximum distance between any centers σ number of centers d max m Then the activation function of hidden neuron becomes: i ϕ i m ( ) x ti exp x t i d max 5 Rossella Cancelliere 6
Learning Algorithm Weights: are computed by means of the pseudo-inverse method. For an example ( x i, d ) consider the output of i the network y( xi ) w ϕ( xi t ) +... + wm ϕ m ( xi tm ) We would like y ( x i ) d i for each example, that is w ϕ ( xi t ) +... + wmϕ m( xi tm ) d i 5 3 Learning Algorithm This can be re-written in matrix form for one example and T [ ϕ ( x i t )... ϕm ( xi tm ) ][ w... wm ] di ϕ( x t )... ϕm ( x tm )... [ w... wm] ϕ( x t )... ϕm ( x tm ) for all the examples at the same time T [ d... d ] T 5 4 Rossella Cancelliere 7
Learning Algorithm let then we can write If Φ + ϕ( x t )... Φ... ϕ( x t )... w Φ... w ϕ ( x t t ) ) is the pseudo-inverse of the matrix we obtain the weights using the following formula T + [... wm] Φ [ d m m m m w... d ] ϕ ( x d... m d Φ 5 5 T Learning Algorithm : summary. Choose the centers randomly from the training set.. Compute the spread for the function using the normalization method. 3. Find the weights using the pseudo-inverse method. 5 6 Rossella Cancelliere 8
Learning Algorithm : Centers clustering algorithm for finding the centers Initialization: t k (0) random k,, m Sampling: draw x from input space 3 Similarity matching: find index of center closer to x k(x) arg min 4 Updating: adjust centers t k (n + ) x(n) t (n) 5 Continuation: increment n by, goto and continue until no noticeable changes of centers occur k [ x(n) t (n)] if k k(x) tk (n) +η k t k (n) otherwise 5 7 k Learning Algorithm : summary Hybrid Learning Process: Clustering for finding the centers. Spreads chosen by normalization. LMS algorithm (see Adaline) for finding the weights. 5 8 Rossella Cancelliere 9
Learning Algorithm 3 Apply the gradient descent method for finding centers, spread and weights, by minimizing the (instantaneous) squared error E ( y( x) d Update for: centers spread weights η E t j t j t j σ j η σ j w ij E σ E η ij w ij j ) 5 9 Comparison with FF -etworks are used for regression and for performing complex (non-linear) pattern classification tasks. Comparison between networks and FF: Both are examples of non-linear layered feed-forward networks. Both are universal approximators. 5 0 Rossella Cancelliere 0
Comparison with multilayer Architecture: networks have one single hidden layer. FF networks may have more hidden layers. euron Model: In the neuron model of the hidden neurons is different from the one of the output nodes. Typically in FF hidden and output neurons share a common neuron model. The hidden layer of is non-linear, the output layer of is linear. Hidden and output layers of FF are usually non-linear. 5 Comparison with multilayer Activation functions: The argument of activation function of each hidden neuron in a computes the Euclidean distance between input vector and the center of that unit. The argument of the activation function of each hidden neuron in a FF computes the inner product of input vector and the synaptic weight vector of that neuron. Approximation: using Gaussian functions construct local approximations to non-linear I/O mapping. FF construct global approximations to non-linear I/O mapping. 5 Rossella Cancelliere
Example: Astronomical image processing We used neural networks to verify the similarity of real astronomical images to predefined reference profiles, so we analyse realistic images and deduce from them a set of parameters able to describe their discrepancy with respect to the ideal, non-aberrated image. The focal plane image, in an ideal case, is the Airy function (blue line): I () r P S λ R p J πr ( πr ) Usually the image is perturbed by aberrations associated to the characteristics of real-world instruments (red line represents the perturbed image) 5 3 Example: Astronomical image processing A realistic image is associated to the Fourier transform of the aperture function: I PS π p iφ( ρ, θ ) iπrρ cos( θ φ ) ( r, φ) dρ dθ ρ e e π λ R 0 0 which includes the contributions of the classical (Seidel) aberrations: Φ π 4 3 ( ρ, θ ) [ A ρ + A ρ cosθ + A ρ cos θ + A ρ + A ρ cosθ ] λ A s : Spherical aberration s c a A c : Coma Aa d : Astigmatism t A d : Field curvature (d defocus) A t : Distortion (t tilt) 5 4 Rossella Cancelliere
Example: Astronomical image processing - The maximum range of variation considered is ± 0.3λ for the training set and ± 0.5λ for the test set - To ease the computational task the image is encoded using the first moments:,0 0, 0, 3 0, 4,,, 3,,, 3 where: nm x y x y dx dy σ x σ y dx dy I ( x, y) n m I ( x, y) 5 5 Example: Astronomical image processing We performed this task using a F neural network (with 60 hidden nodes) and a neural network (with 75 hidden nodes) Usually the former requires less internal nodes, but more training iterations than the latter The training required 000 iterations (F) and 75 iterations (); both optimised networks provide reasonably good results in terms of approximation of the desired unknown function relating aberrations and moments. 5 6 Rossella Cancelliere 3
Example: Astronomical image processing Performances of the sigmoidal (left) and radial (right) neural networks; from top to bottom, coma, astigmatism, defocus and distortion are shown. The blue (solid) line follows the test set targets, whereas the green dots are the values computed by the networks. 5 7 Rossella Cancelliere 4