Showing papers on "Convolutional neural network published in 1999"

Object Recognition with Gradient-Based Learning

Abstract: Finding an appropriate set of features is an essential problem in the design of shape recognition systems. This paper attempts to show that for recognizing simple objects with high shape variability such as handwritten characters, it is possible, and even advantageous, to feed the system directly with minimally processed images and to rely on learning to extract the right set of features. Convolutional Neural Networks are shown to be particularly well suited to this task. We also show that these networks can be used to recognize multiple objects without requiring explicit segmentation of the objects from their surrounding. The second part of the paper presents the Graph Transformer Network model which extends the applicability of gradient-based learning to systems that use graphs to represents features, objects, and their combinations.

Combined classifiers for invariant face recognition

Abstract: We present a system for invariant face recognition. A combined classifier uses the generalization capabilities of both learning vector quantization (LVQ) and radial basis function (RBF) neural networks to build a representative model of a face from a variety of training patterns with different poses, details and facial expressions. The combined generalization error of the classifier is found to be lower than that of each individual classifier. A new face synthesis method is implemented for reducing the false acceptance rate and enhancing the rejection capability of the classifier. The system is capable of recognizing a face in less than one second. The system is tested on the well-known ORL database. The system performance compares favorably with the state-of-the-art systems. In the case of the ORL database, a correct recognition rate of 99.5% at 0.5% rejection rate is achieved. This rate compares favorably with the rates achieved by other systems on the same database. The volumetric frequency domain representation resulted in a rate of 92.5% while the combination of a convolutional neural network and self-organizing map resulted in 96.2% for the same number of training faces (five) per person in a database representing 40 people.

Topographic Transformation as a Discrete Latent Variable

Abstract: Invariance to topographic transformations such as translation and shearing in an image has been successfully incorporated into feed-forward mechanisms, >e.g., "convolutional neural networks", "tangent propagation". We describe a way to add transformation invariance to a generative density model by approximating the nonlinear transformation manifold by a discrete set of transformations. An EM algorithm for the original model can be extended to the new model by computing expectations over the set of transformations. We show how to add a discrete transformation variable to Gaussian mixture modeling, factor analysis and mixtures of factor analysis. We give results on filtering microscopy images, face and facial pose clustering, and handwritten digit modeling and recognition.

Building pattern classifiers using convolutional neural networks

Abstract: Pattern classification is the core task of many applications such as image segmentation. This paper studies the possibility of building pattern classifiers for text/picture segmentation and text detection problems using convolutional neural networks (CNNs). By using CNNs, explicit feature extraction is avoided-the feature detectors are learned from the training data. More importantly, CNNs can directly operate on grey level images, making its application straightforward. Addressed are practical issues such as kernel size, convergence speed, etc. Experiments on Chinese text/picture segmentation and text detection are presented.

Learning in intelligent embedded systems

Abstract: Information processing capabilities of embedded systems presently lack the robustness and rich complexity found in biological systems. Endowing artificial systems with the ability to adapt to changing conditions requires algorithms that can rapidly learn from examples. We demonstrate the application of one such learning algorithm on an inexpensive robot constructed to perform simple sensorimotor tasks. The robot learns to track a particular object by discovering the salient visual and auditory cues unique to that object. The system uses a convolutional neural network to combine color, luminance, motion, and auditory information. The weights of the networks are adjusted using feedback from a teacher to reflect the reliability of the various input channels in the surrounding environment. We also discuss how unsupervised learning can discover features in data without external interaction. An unsupervised algorithm based upon nonnegative matrix factorization is able to automatically learn the different parts of objects. Such a parts-based representation of data is crucial for robust object recognition.

A dynamic neural network for syllable recognition

Abstract: A dynamic neural network architecture based on the time-delay neural network and the convolutional neural network is originated. The dynamic network achieves much better performance than those of MLP and TDNN when dealing with syllable recognition. Such performance is also comparable to that of the more popular HMM method.

On the overfitting of the five-layered bottleneck network

Abstract: In autoassociative learning for the bottleneck neural network, the problem of overfitting is pointed out. This overfitting is pathological in the sense that it does not disappear even if the sample size goes to infinity. However, it is not observed in the real learning process. Thus we study the basin of the overfitting solution. First, the existence of overfitting is confirmed. Then it is shown that the basin of the overfitting solution is small compared with the normal solution.

A genetic algorithm approach used to generate the neural network structures

Abstract: A genetic algorithm (GA) is implemented to search for the optimal structures of neural networks which are used for approximating a given nonlinear function. Two kinds of neural networks, i.e. the multilayer feedforward and time delay neural networks are involved in the paper. The weights of each neural network in each generation are obtained by associated training algorithms. The simulation results of nonlinear function approximation are given and some improvements in the future are outlined.

Neural net computing for biomedical image processing

Abstract: In this paper we describe some of the most important types of neural networks applied in biomedical image processing. The networks described are variations of well-known architectures but are including image-relevant features in their structure. Convolutional neural networks, modified Hopfield networks, regularization networks and nonlinear principal component analysis neural networks are successfully applied in biomedical image classification, restoration and compression.