A Connection Between GRBF and MLP

Both multilayer perceptrons (MLP) and Generalized Radial Basis Functions (GRBF) have good approximation properties, theoretically and experimentally. Are they related? The main point of this paper is to show that for normalized inputs, multilayer perceptron networks are radial function networks (albeit with a non-standard radial function). This provides an interpretation of the weights w as centers t of the radial function network, and therefore as equivalent to templates. This insight may be useful for practical applications, including better initialization procedures for MLP. In the remainder of the paper, we discuss the relation between the radial functions that correspond to the sigmoid for normalized inputs and well-behaved radial basis functions, such as the Gaussian. In particular, we observe that the radial function associated with the sigmoid is an activation function that is good approximation to Gaussian basis functions for a range of values of the bias parameter. The implication is that a MLP network can always simulate a Gaussian GRBF network (with the same number of units but less parameters); the converse is true only for certain values of the bias parameter. Numerical experiments indicate that this constraint is not always satisfied in practice by MLP networks trained with backpropagation. Multiscale GRBF networks, on the other hand, can approximate MLP networks with a similar number of parameters.

Linear Object Classes and Image Synthesis from a Single Example Image

The need to generate new views of a 3D object from a single real image arises in several fields, including graphics and object recognition. While the traditional approach relies on the use of 3D models, we have recently introduced techniques that are applicable under restricted conditions but simpler. The approach exploits image transformations that are specific to the relevant object class and learnable from example views of other "prototypical" objects of the same class. In this paper, we introduce such a new technique by extending the notion of linear class first proposed by Poggio and Vetter. For linear object classes it is shown that linear transformations can be learned exactly from a basis set of 2D prototypical views. We demonstrate the approach on artificial objects and then show preliminary evidence that the technique can effectively "rotate" high- resolution face images from a single 2D view.

RABBIT: A Compiler for SCHEME

We have developed a compiler for the lexically-scoped dialect of LISP known as SCHEME. The compiler knows relatively little about specific data manipulation primitives such as arithmetic operators, but concentrates on general issues of environment and control. Rather than having specialized knowledge about a large variety of control and environment constructs, the compiler handles only a small basis set which reflects the semantics of lambda-calculus. All of the traditional imperative constructs, such as sequencing, assignment, looping, GOTO, as well as many standard LISP constructs such as AND, OR, and COND, are expressed in macros in terms of the applicative basis set. A small number of optimization techniques, coupled with the treatment of function calls as GOTO statements, serve to produce code as good as that produced by more traditional compilers. The macro approach enables speedy implementation of new constructs as desired without sacrificing efficiency in the generated code. A fair amount of analysis is devoted to determining whether environments may be stack-allocated or must be heap-allocated. Heap-allocated environments are necessary in general because SCHEME (unlike Algol 60 and Algol 68, for example) allows procedures with free lexically scoped variables to be returned as the values of other procedures; the Algol stack-allocation environment strategy does not suffice. The methods used here indicate that a heap-allocating generalization of the "display" technique leads to an efficient implementation of such "upward funargs". Moreover, compile-time optimization and analysis can eliminate many "funargs" entirely, and so far fewer environment structures need be allocated at run time than might be expected. A subset of SCHEME (rather than triples, for example) serves as the representation intermediate between the optimized SCHEME code and the final output code; code is expressed in this subset in the so-called continuation-passing style. As a subset of SCHEME, it enjoys the same theoretical properties; one could even apply the same optimizer used on the input code to the intermediate code. However, the subset is so chosen that all temporary quantities are made manifest as variables, and no control stack is needed to evaluate it. As a result, this apparently applicative representation admits an imperative interpretation which permits easy transcription to final imperative machine code. These qualities suggest that an applicative language like SCHEME is a better candidate for an UNCOL than the more imperative candidates proposed to date.

Comparing Support Vector Machines with Gaussian Kernels to Radial Basis Function Classifiers

The Support Vector (SV) machine is a novel type of learning machine, based on statistical learning theory, which contains polynomial classifiers, neural networks, and radial basis function (RBF) networks as special cases. In the RBF case, the SV algorithm automatically determines centers, weights and threshold such as to minimize an upper bound on the expected test error. The present study is devoted to an experimental comparison of these machines with a classical approach, where the centers are determined by $k$--means clustering and the weights are found using error backpropagation. We consider three machines, namely a classical RBF machine, an SV machine with Gaussian kernel, and a hybrid system with the centers determined by the SV method and the weights trained by error backpropagation. Our results show that on the US postal service database of handwritten digits, the SV machine achieves the highest test accuracy, followed by the hybrid approach. The SV approach is thus not only theoretically well--founded, but also superior in a practical application.

Priors Stabilizers and Basis Functions: From Regularization to Radial, Tensor and Additive Splines

We had previously shown that regularization principles lead to approximation schemes, as Radial Basis Functions, which are equivalent to networks with one layer of hidden units, called Regularization Networks. In this paper we show that regularization networks encompass a much broader range of approximation schemes, including many of the popular general additive models, Breiman's hinge functions and some forms of Projection Pursuit Regression. In the probabilistic interpretation of regularization, the different classes of basis functions correspond to different classes of prior probabilities on the approximating function spaces, and therefore to different types of smoothness assumptions. In the final part of the paper, we also show a relation between activation functions of the Gaussian and sigmoidal type.