One ofthe major theoretical issues in neural computation is to understand how this efficiency is reached given the constraints imposed by the biological hardware. Part of the problem [2] is simply to give an informative representation ofthe visual world using a limited number of neurons, each of which has a limited information capacity. The information capacity of the visual system is determined in part by the spatial transfer characteristics, or "receptive fields," of the individual cells. From a theoretical point of view we can ask if there exists an optimal choice for these receptive fields, a choice which maximizes the information transfer through the system given the hardware constraints. We show that this optimization problem has a simple formulation which allows us to use the intuition developed through the variational approach to quantum mechanics. In general our approach leads to receptive fields which are quite unlike those observed forcells in the visual cortex. In particular orientation selectivity is not a generic prediction. The optimal filters, however, depend on the statistical properties ofthe images we are trying to sample. Natural images have a symmetry - scale invariance [4] - which saves the theory: The optimal receptive fields for sampling of natural images are indeed orientation selective and bear a striking resemblance to observed receptive field characteristics in the mammalian visual cortex as well as the retinal ganglion of lower vertebrates.

A neural network architecture was designed for locating word boundaries and identifying words from phoneme sequences. This architecture was tested in of studies.

We compare the performance of the modular architecture, composed of competing expert networks, suggested by Jacobs, Jordan, Nowlan and Hinton (1991) to the performance of a single back-propagation network on a complex, but low-dimensional, vowel recognition task. Simulations reveal that this system is capable of uncovering interesting decompositions in a complex task. The type of decomposition is strongly influenced by the nature of the input to the gating network that decides which expert to use for each case. The modular architecture also exhibits consistently better generalization on many variations of the task. 1 Introduction If back-propagation is used to train a single, multilayer network to perform different subtasks on different occasions, there will generally be strong interference effects which lead to slow learning and poor generalization. If we know in advance that a set of training cases may be naturally divideJ into subsets that correspond to distinct subtasks, interference can be reduced by using a system (see Figure 1) composed of several different "expert" networks plus a gating network that decides which of the experts should be used for each training case. Systems of this type have been suggested by a number of authors (Hampshire and Waibel, 1989; Jacobs, Jordan and Barto, 1990; Jacobs et al., 1991) (see also the paper by Jacobs and Jordan in this volume (1991».

A novel unsupervised neural network for dimensionality reduction which seeks directions emphasizing multimodality is presented, and its connection toexploratory projection pursuit methods is discussed. This leads to a new statistical insight to the synaptic modification equations governing learning in Bienenstock, Cooper, and Munro (BCM) neurons (1982). The importance of a dimensionality reduction principle based solely on distinguishing features, is demonstrated using a linguistically motivated phoneme recognition experiment, and compared with feature extraction using back-propagation network. 1 Introduction Due to the curse of dimensionality (Bellman, 1961) it is desirable to extract features froma high dimensional data space before attempting a classification. How to perform this feature extraction/dimensionality reduction is not that clear. A first simplification is to consider only features defined by linear (or semi-linear) projections ofhigh dimensional data.

For lack of alternative models, search and decision processes have provided the dominant paradigm for human memory access using two or more cues, despite evidence against search as an access process (Humphreys, Wiles & Bain, 1990). We present an alternative process to search, based on calculating the intersection of sets of targets activated by two or more cues. Two methods of computing the intersection are presented, one using information about the possible targets, the other constraining the cue-target strengths in the memory matrix. Analysis using orthogonal vectors to represent the cues and targets demonstrates the competence of both processes, and simulations using sparse distributed representations demonstrate the performance of the latter process for tasks involving 2 and 3 cues.

In this paper, we will describe several extensions to our earlier work, utilizing asegment-based approach. We will formulate our segmental framework and report our study on the use of multi-layer perceptrons for detection and classification of phonemes. We will also examine the outputs of the network, and compare the network performance with other classifiers. Our investigation is performed within a set of experiments that attempts to recognize 38 vowels and consonants in American English independent of speaker.

In this paper we show that discrete affine wavelet transforms can provide a tool for the analysis and synthesis of standard feedforward neural networks. Itis shown that wavelet frames for L2(IR) can be constructed based upon sigmoids. The spatia-spectral localization property of wavelets can be exploited in defining the topology and determining the weights of a feedforward network. Training a network constructed using the synthesis proceduredescribed here involves minimization of a convex cost functional andtherefore avoids pitfalls inherent in standard backpropagation algorithms. Extension of these methods to L2(IRN) is also discussed. 1 INTRODUCTION Feedforward type neural network models constructed from empirical data have been found to display significant predictive power [6]. Mathematical justification in support ofsuch predictive power may be drawn from various density and approximation theorems [1, 2, 5].

Van den Bout North Carolina State University Box 7914 Raleigh, NC 27695-7914 We apply the theory of Tishby, Levin, and Sol1a (TLS) to two problems. First we analyze an elementary problem for which we find the predictions consistent with conventional statistical results. Second we numerically examine the more realistic problem of training a competitive net to learn a probability density from samples. We find TLS useful for predicting average training behavior. . 1 TLS APPLIED TO LEARNING DENSITIES Recently a theory of learning has been constructed which describes the learning of a relation from examples (Tishby, Levin, and Sol1a, 1989), (Schwarb, Samalan, Sol1a, and Denker, 1990). The original derivation relies on a statistical mechanics treatment of the probability of independent events in a system with a specified average value of an additive error function. The resulting theory is not restricted to learning relations and it is not essentially statistical mechanical. The TLS theory can be derived from the principle of mazimum entropy,a general inference tool which produces probabilities characterized by certain values of the averages of specified functions(Jaynes, 1979). A TLS theory can be constructed whenever the specified function is additive and associated with independent examples. In this paper we treat the problem of learning a probability density from samples.

Dean A. Pomerleau School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213 Abstract The ALVINN (Autonomous Land Vehicle In a Neural Network) project addresses the problem of training artificial neural networks in real time to perform difficult perception tasks. ALVINN,is a back-propagation network that uses inputs from a video camera and an imaging laser rangefinder to drive the CMU Navlab, a modified Chevy van. This paper describes training techniques which allow ALVINN to learn in under 5 minutes to autonomously control the Navlab by watching a human driver's response to new situations. Using these techniques, ALVINN has been trained to drive in a variety of circumstances including single-lane paved and unpaved roads, multilane lined and unlined roads, and obstacle-ridden on-and off-road environments, at speeds of up to 20 miles per hour. 1 INTRODUCTION Previous trainable connectionist perception systems have often ignored important aspects of the form and content of available sensor data. Because of the assumed impracticality of training networks to perform realistic high level perception tasks, connectionist researchers have frequently restricted their task domains to either toy problems (e.g. the TC identification problem [11] [6]) or fixed low level operations (e.g.

This paper extends work reported in (Hayashi & Nakai.