Akito Sakurai

Using digital libraries as a community hall for worldwide information spiral development

Our concept of a digital library from a users viewpoint is explained by using a three-layer model of information distribution: information provider, information broker, and information user. Our worldwide communication-hall prototype consists of the virtual personal library client, called Webshelf, an information publishing server, and a hypermedia archive server, Webarchive. It allows users to virtually personalize information on the Web, organize it, create new information, and publish this new information. Webarchive offers an effective solution to the information volatility problem by storing all of the information obtained by users and managing link consistency.

n-h-1 networks store no less n*h+1 examples, but sometimes no more

The author shows that an n-h-1 artificial neural network with n real inputs, a single layer of h hidden units, and one binary output unit can store correctly at least n*h+1 examples in a general position. The proof is constructive so that weights are obtained deterministically from examples. The result is thought to be a generalization of the fact that one threshold gate can remember any n+1 examples in a general position. The number obtained is a good lower bound of the network capacity and is a great improvement on the previous best bound by S. Akaho and S. Amari (1990). It is also shown that the figure nh+1 is tight in a certain sense.<<ETX>>

Providing definite clauses without explicit use of inductions

We propose an algorithm to prove a definite clause is modeled by an intended minimal model of a set of definite clauses. In the algorithm, no induction scheme will be used explicitly, although in effect a mathematical induction is implied. This algorithm corresponds to an inductionless induction method using the Knuth-Bendix algorithm in an equation problem. It resembles the completion algorithm, but in fact it is a completely different algorithm.

On the VC-dimension of depth four threshold circuits and the complexity of Boolean-valued functions

We consider the problem of determining VC-dimension ∂3(h) of depth four n-input 1-output threshold circuits with h elements. Best known asymptotic lower bounds and upper bounds are proved, that is, when h → ∞, ∂3(h) is upper bounded by ((h2/3)+nh(log h)(1+o(1))) and lower bounded by (1/2)((h2/4)+nh)(log h)(1 − o(1)). We also consider the problem of determining complexity c3(N) of Boolean-valued functions defined on N-pointsets in Rn, measured by the number of threshold elements, with which we can construct a depth four circuit to realize the functions. We also show the best known upper and lower bounds, that is, when N → ∞, the complexity is upper bounded by(sqrt {16left( {{N mathord{left/{vphantom {N {log N}}} right.kern-nulldelimiterspace} {log N}}} right)left( {1 + o(1)} right) + 4n^2 } - 2n) and lower bounded by (sqrt {6left( {{N mathord{left/{vphantom {N {log N}}} right.kern-nulldelimiterspace} {log N}}} right)left( {1 + o(1)} right)left( {{9 mathord{left/{vphantom {9 4}} right.kern-nulldelimiterspace} 4}} right)n^2 } - left( {{3 mathord{left/{vphantom {3 2}} right.kern-nulldelimiterspace} 2}} right)n)

Shape discrimination using integral features

Integral features are calculated by summing the local features over all the pixels, where the local features are determined by the state of the neighborhood of each pixel. The states are defined by using a two-dimensional series of mask functions on the polar coordinate system with the logarithmic scale of r-direction. This definition enables the efficient extraction of features from any arbitrary distant area. The features for shape discrimination are constructed from the short-range correlations of the gradients of the image data. For discrimination of image data we used the linear model as used in the multivariate analysis. We also developed nonlinear model learning by maximizing the discriminant efficiency. In the models, each pixel has the value that represents the validity of discrimination and weighted summations are performed when the integral features are calculated. The validity of the linear and nonlinear models is verified in experiments using the image data of real objects.