Algebraic transformations of objective functions
نویسندگان
چکیده
Many neural networks can be derived as optimization dynamics for suitable objective functions. We show that such networks can be designed by repeated transformations of one objective into another with the same xpoints. We exhibit a collection of algebraic transformations which reduce network cost and increase the set of objective functions that are neurally implementable. The transformations include simpliication of products of expressions, functions of one or two expressions, and sparse matrix products (all of which may be interpreted as Legendre transformations); also the minimum and maximum of a set of expressions. These transformations introduce new interneurons which force the network to seek a saddle point rather than a minimum. Other transformations allow control of the network dynamics, by reconciling the Lagrangian formalism with the need for xpoints. We apply the transformations to simplify a number of structured neural networks, beginning with the standard reduction of the winner-take-all network from O(N 2) connections to O(N). Also susceptible are inexact graph-matching, random dot matching, convolutions and coordinate transformations, and sorting. Simulations show that xpoint-preserving transformations may be applied repeatedly and elaborately, and the example networks still robustly converge.
منابع مشابه
Transformations of hypergeometric elliptic integrals
The paper classifies algebraic transformations of Gauss hypergeometric functions with the local exponent differences (1/2, 1/4, 1/4), (1/2, 1/3, 1/6) and (1/3, 1/3, 1/3). These form a special class of algebraic transformations of Gauss hypergeometric functions, of arbitrary high degree. The Gauss hypergeometric functions can be identified as elliptic integrals on the genus 1 curves y = x − x or...
متن کاملContinuity of super- and sub-additive transformations of continuous functions
We prove a continuity inheritance property for super- and sub-additive transformations of non-negative continuous multivariate functions defined on the domain of all non-negative points and vanishing at the origin. As a corollary of this result we obtain that super- and sub-additive transformations of continuous aggregation functions are again continuous aggregation functions.
متن کاملAlgebraic transformations of Gauss hypergeometric functions
This paper classifies algebraic transformations of Gauss hypergeometric functions and pull-back transformations between hypergeometric differential equations. This classification recovers the classical transformations of degree 2, 3, 4, 6, and finds other transformations of some special classes of the Gauss hypergeometric function.
متن کاملAdmissible transformations and assignment problems
We introduce the notion of admissible transfomations which is related to the Hungarian method for solving assignment problems. Admissible transformations are stated for linear, quadratic and multi-index assignment problems. Their application to find good lower bounds and/or to solve the problem, respectively, is outlined. Finally it is shown that admissible transformations can also be applied t...
متن کاملذخیره در منابع من
با ذخیره ی این منبع در منابع من، دسترسی به آن را برای استفاده های بعدی آسان تر کنید
برای دانلود متن کامل این مقاله و بیش از 32 میلیون مقاله دیگر ابتدا ثبت نام کنید
ثبت ناماگر عضو سایت هستید لطفا وارد حساب کاربری خود شوید
ورودعنوان ژورنال:
- Neural Networks
دوره 3 شماره
صفحات -
تاریخ انتشار 1990