This paper describes how new learning methods may make it possible for a large‐scale, hierarchical neural network to recognize most Japanese handwritten characters. This is a very large and complex task, as the Japanese character set consists of about 3000 categories which can be written in many different ways. Such a difficult task can lead a neural network to converge very slowly and to yield recognition rates that are uneven between categories. To address these problems we here propose five learning methods as modifications of the conventional back‐propagation learning rule. These methods produce fast convergence, even recognition rates over all categories, and adequate recognition of test samples. We also describe how a large‐scale neural network can be built by dividing the recognition task into several subtasks, with networks for each subtask, and then integrating these subnetworks in a large network with a hierarchical structure. In a hierarchical network, the upper level network directly integrates outputs from each lower level network. Application of that network to handwritten Japanese character recognition has resulted in poor recognition, because lower level networks do not know about unknown input patterns, and the direct integration of ambiguous outputs from many lower level networks confuses the upper level network. We propose a new integration method which provides each subnetwork with more information as to how close an input pattern is to the categories of that subnetwork. This method resulted in high recognition performance for character recognition. We here described the above methods, and report the performance of our implementation of a neural network for the recognition of 71 Hiragana characters, and describe our implementation of this network on a hypercube concurrent computer.

No abstract

SUMMARYThis paper investigates suitable indexing techniques to enable efficient content-based audio retrieval in large acoustic databases. To make an index-based retrieval mechanism applicable to audio content, we investigate the design of Locality Sensitive Hashing (LSH) and the partial sequence comparison. We propose a fast and efficient audio retrieval framework of query-by-content and develop an audio retrieval system. Based on this framework, four different audio retrieval schemes, LSH-Dynamic Programming (DP), LSH-Sparse DP (SDP), Exact Euclidian LSH (E2LSH)-DP, E2LSH-SDP, are introduced and evaluated in order to better understand the performance of audio retrieval algorithms. The experimental results indicate that compared with the traditional DP and the other three compititive schemes, E2LSH-SDP exhibits the best tradeoff in terms of the response time, retrieval accuracy and computation cost. key words: indexing, locality-sensitive hashing, content-based audio retrieval, dynamic programming 1. Introduction Content-based audio retrieval is not only a very promising research topic but also one of the main problems, in multimedia information processing. Handling audio sequence data is usually time-consuming due to the high dimensionality of the features, which makes it inconvenient to utilize the potential content-based information retrieval techniques on the Internet or personal media devices. To access a huge mass of audio information efficiently, it is necessary to explore the audio information, facilitate the management of audio data and serve multimedia applications. Consequently various indexing structures have been reported in the study of audio retrieval. These include, for example, hierarchical structure As far as the creation of query-by-content audio retrieval mechanism via indexing techniques is concerned the main challenges are as follows: (1)

No abstract

Sparse matrix-vector multiplication on GPUs faces to a serious problem when the vector length is too large to be stored in GPU's device memory. To solve this problem, we propose a novel software-hardware hybrid method for a heterogeneous system with GPUs and functional memory modules connected by PCI express. The functional memory contains huge capacity of memory and provides scatter/gather operations. We perform some preliminary evaluation for the proposed method with using a sparse matrix benchmark collection. We observe that the proposed method for a GPU with converting indirect references to direct references without exhausting GPU's cache memory achieves 4.1 times speedup compared with conventional methods. The proposed method intrinsically has high scalability of the number of GPUs because intercommunication among GPUs is completely eliminated. Therefore we estimate the performance of our proposed method would be expressed as the single GPU execution performance, which may be suppressed by the burst-transfer bandwidth of PCI express, multiplied with the number of GPUs.