Recently, performance of deep neural networks, especially convolutional neural networks (CNNs), has been drastically increased by elaborate network architectures, by new learning methods, and by GPU-based high-performance computation. However, there are still several difficult problems concerning back propagation, which include scheduling of learning rate and controlling locality of search (i.e., avoidance of bad local minima). A learning method, called “learning-rate-optimizing genetic back-propagation” (LOG-BP), which combines back propagation with a genetic algorithm by a new manner, is proposed. This method solves the above-mentioned two problems by optimizing the learning process, especially learning rate, by genetic mutations and by locality-controlled parallel search. Initial experimental results shows that LOG-BP performs better; that is, when required, learning rate decreases exponentially and the distances between chromosomes, which indicate the locality of a search, also decrease exponentially.

We introduce the latest extended functions for the VNode infrastructure. We present new extended VNode infrastructure functions that achieve high performance and provide convenient deep programmability to network developers. In addition, we extend network virtualization from the core network to edge networks and terminals. We deploy an enhanced VNode infrastructure on the JGN-X testbed in evaluation experiments. We also succeeded to create international federation slice between GENI and Fed4FIRE.

An architecture called the slice-exchange-point (SEP) has been designed for federating heterogeneous net-work-virtualization platforms by creating and managing slices (virtual networks). SEP enables whole inter-domain resources to be managed by the network manager of any single domain. Slice-operation commands are propagated to other domains through SEP by using a common API. SEP introduces the following four features: infrastructure neutrality, single interface federation, abstract and clean federation, and extensibility of capabilities. SEP's functions to achieve these features are discussed. SEP was partially implemented on two VNode domains and one ProtoGENI domain and was verified to function effectively.

Network virtualization introduces two concepts: slice (i.e., virtual network), which consists of virtual nodes and links, and slice developer, which is the third role in networks. Slice developers can introduce new network services by using slices. A method for introducing new types of virtual nodes and links for new services into the slice-definition language and the virtualization infrastructure by evolving physical nodes (i.e., “virtualization nodes” or VNodes) is proposed. This evolution consists of two stages: the experimental stage and the operational stage. In the experimental stage, data and control plug-ins are developed and tested by the operator or vendor by using experimental sliver definitions. In the operational stage, which is focused on in this study, the plug-ins are integrated into original components in the infrastructure and are available for slice development by using normal sliver definitions. By mapping type names to plug-in identifiers and parameters, the proposed method enables abstract and simple definitions of slices by slice developers and authorization of plug-ins by the operator, but it remains the loose integration of the new function, i.e., the plug-in architecture used in the experimental stage. Prototyping and evaluation demonstrates that this method greatly simplifies both slice developers' tasks and operators' tasks.

One key requirement for achieving network virtualization is resource isolation among slices (virtual networks), that is, to avoid interferences between slices of resources. This paper proposes two methods, per-slice shaping and per-link policing for network-resource isolation (NRI) in terms of bandwidth and delay. These methods use traffic shaping and traffic policing, which are widely-used traffic control methods for guaranteeing QoS. Per-slice shaping utilizes weighted fair queuing (WFQ) usually applied to a fine grained flow such as a flow from a specific server application to a user. Since the WFQ for fine-grained flows requires many queues, it may not scale to a large number of slices with a large number of virtual nodes. Considering that the purpose of NRI is not thoroughly guaranteeing QoS but avoiding interferences between slices, we believe per-slice shaping suffices our objective. In contrast, per-link policing uses traffic policing per virtual link. It requires less resource and achieves less strict isolation between hundreds of slices. Our results show that both methods perform NRI well but the performance of the former is better in terms of delay. Accordingly, per-slice shaping is effective for delay-sensitive services while per-link policing may be sufficiently used for the other types of services.

We are developing a new voice communication medium called voiscape. Voiscape enables natural and seamless bi-directional voice communication by using sound to create a virtual sound room. In a sound room, people can feel others' direction and distance expressed by spatial sounds with reverberations, and they can move freely by using a map of the room. Voiscape enables multi-voice-conversations. In a virtual market place that will be realized by voiscape, people can not only buy goods or information but also enjoy talking with merchants and people there. In this demo, a voiscape prototype called VPII is used for realizing such an environment. Unfortunately, because prerecorded voices are used in this demo, the participants cannot talk with merchants. However, the participants can talk each other with small end-to-end latency (less than 200 ms) and will feel the atmosphere of the virtual market place. Prerecorded people and merchants talk each other in English, Japanese and Chinese in parallel and with crossovers, and participants can virtually walk among them and can selectively listen one voice or hear multiple voices at once.

An improved prototype of the "voiscape" voice communication medium has been developed and subjectively evaluated. Voiscape enables natural and seamless voice communication by using sound to create a virtual "sound room" in which people, who are represented by different sounds, can move freely. It features low-delay motion-tracking spatial audio with simulated early reflections that produce out-of-head sound localization and sound distance expression. It also features virtual-location-based selective communication: a user can walk freely in the sound room using a map- and cursor-key-based user-interface and can select whom to talk to or which sound sources to listen to. A third feature is SIP-presence-event-notification (SIMPLE)-based sound room management: when users move, their locations and directions are distributed using SIP SUBSCRIBE/NOTIFY messages. The combination of these features creates a natural voice-communication space in which two or more parallel conversation contexts can coexist. Limited, subjective testing by around 200 people showed that this medium can be used for cocktail-party-like conversation; i.e., users could distinguish parallel conversations by paying attention to or by moving toward one of them.

Because higher- and lower-level policies do not necessarily correspond one-to-one, a higher-level network policy may have to be translated into two or more lower-level policies, and two or more cooperating higher-level policies may have to be translated into one lower-level policy. The former transformation is called a policy division, and the latter transformation is called a policy fusion. These transformations can be performed mechanically under certain restricted conditions, as described in this paper. In general, however, such transformations are very complicated, and the restrictions cannot be eliminated completely. This is mainly due to the existence of multiple packet classifiers in a set of policies. This paper thus concludes that multiple classifiers should not be introduced, if possible. Policy division and fusion can be avoided in certain cases, but are probably unavoidable in other cases. Given this, these problems should be solved by introducing virtual flow labels to remove harmful classifiers and conducting further studies on policy division and fusion.

A text retrieval method called the thematic geographical search method has been developed and applied to a Japanese encyclopedia called the World Encyclopædia. In this method, the user specifies a search theme using free words, then obtains a sorted list of excerpts and hyperlinks to encyclopedia sentences that contain geographical names. Using this list, the user can also open maps that indicate the locations of the names. To generate an index of names for this searching, a method of extracting geographical names has been developed. In this method, geographical names are extracted, matched to names in a geographical name database, and identified. Geographical names, however, often have several types of ambiguities. Ambiguities are resolved by using non-local context analysis, which uses a stack and several other techniques. As a result, the precision of extracted names is more than 96% on average. This method depends on features of the Japanese language, but the strategy and most of the techniques can be applied to texts in English or other languages.

The conventional processing techniques for pipelined vector processors such as Cray-XMP, or SIMD parallel processors, such as CM-2 {connection machine), are generally applied only to independent multiple data processing.This paper describes a vector processing method of multiple processing including parallel rewriting of dynanu"c data structures with shared elements, and of multiple processing that may rewrite the same data element two or more times.This method is called the jVteringoverwritten-label method (FOL).FOL enables vector processing of entering multiple data into a hash table, address calculation sorting, and many other algorithms that handle lists, trees, graphs and other types of symbolic data structures.FOL is applied to several symbolic processing algorithms; consequently, the pe~ormance is improved by a factor often on the Hitachi S-81O.

The problems of utilizing an extended vector processor called IDP (Integrated Database Processor) for nonnumerical processing are discussed. IDP has highly pipelined merge and search vector facilities in addition to conventional vector facilities. This extended vector processor can be utilized not only for primitive relational operations but also for the following applications: (1) complex relational query using multiple indexes; (2) multiway merge/sort utility, and (3) generation and test processing. Vectorization methods and performance improvements are described and analyzed.>

Several techniques for running Prolog programs on pipelined vector processors, such as the Hitachi S-820 or the Cray-2, are developed. This paper presents an automatic program transformation (vectorization) method of Prolog, which enables a type of or-parallel execution of Prolog programs using vector operations. Performance is evaluated on the Hitachi S-810 using the Eight-Queens problem. Its vector execution speed is 4.5 MLIPS (18 ms). This is eight or nine times faster than scalar execution. This result confirms the effectiveness of vectorization techniques and applicability of vector processors to Prolog execution and to symbol processing applications.