Taro Tezuka

Kernel Analysis for Estimating the Connectivity of a Network with Event Sequences

Abstract Estimating the connectivity of a network from events observed at each node has many applications. One prominent example is found in neuroscience, where spike trains (sequences of action potentials) are observed at each neuron, but the way in which these neurons are connected is unknown. This paper introduces a novel method for estimating connections between nodes using a similarity measure between sequences of event times. Specifically, a normalized positive definite kernel defined on spike trains was used. The proposed method was evaluated using synthetic and real data, by comparing with methods using transfer entropy and the Victor-Purpura distance. Synthetic data was generated using CERM (Coupled Escape-Rate Model), a model that generates various spike trains. Real data recorded from the visual cortex of an anaesthetized cat was analyzed as well. The results showed that the proposed method provides an effective way of estimating the connectivity of a network when the time sequences of events are the only available information.

Visualization of relationships among historical persons from Japanese historical documents

Digitization of Japanese historical documents has gained much attraction in the field of humanities in Japan recently, and numbers of documents are already available in digitized text format. However, text analysis of these documents has rarely been done mainly due to the lack of natural language processing tools that can handle pre-modern Japanese. In this article, we propose a method to extract and visualize the relationships among persons from Japanese historical documents with an aid of supplementary information such as personal name and place name indices. The goal of the method is to extract dynamics of relationships among historical persons. The method utilizes locational information to obtain latent relationships among persons based on their spatial activities. The proposed method is applied to a Japanese historical chronicle written in the 12th century. Experimental results showed a strong correspondence to the known historical facts, and the results of a user survey completed by researchers of Japanese history demonstrated some potential for the method to serve as a new approach in the fields of humanities. .................................................................................................................................................................................

Parametric Learning of Deep Convolutional Neural Network

Deep neural networks have recently been showing great potential on visual recognition tasks. However, it is also considered difficult to tune its parameters, and it has high training cost. This work focuses on analysis of several learning methods and properties of multinomial logistic regression deep convolutional network. We implemented a scalable deep neural network, compared the efficiency of different methods and how parameters affect the learning process. We propose an efficient method of performing back-propagation with limited kernel functions on GPU and achieved better efficiency. Our conclusions can be applied to train deep networks more efficiently. We achieved the recognition rate of over 0.95 without image preprocessing and fine tuning, within 10 minutes on a single machine.

Spike Train kernels for multiple neuron recordings

There is a growing interest in analyzing multineuron spike trains, which are spike timing data obtained from multiple neurons in the brain. Kernel methods have been successful in clustering and classification of single-neuron spike trains. We extend these methods to multineuron spike trains. Among various possible extensions, the mixture kernel was found to be most effective. The optimum parameter obtained from training this kernel was close to a biologically plausible value, suggesting that our approach is effective for seeking an appropriate model for the activity of a set of neurons.

Transformed Reality - Altering Human Perceptions by Computation

We propose Transformed Reality, a new approach that alters our perception of the world using computation into a form that the user prefers. For a specific example of Transformed Reality, we implemented Anime Glasses, a system that turns a natural scene into the style of an anime, or a cartoon film. The system is built on mobile devices, so the user can walk around and live in the altered environment. To achieve this purpose, methods of image and video processing were tested and compared in experiments.

Multineuron spike train analysis with R-convolution linear combination kernel

A spike train kernel provides an effective way of decoding information represented by a spike train. Some spike train kernels have been extended to multineuron spike trains, which are simultaneously recorded spike trains obtained from multiple neurons. However, most of these multineuron extensions were carried out in a kernel-specific manner. In this paper, a general framework is proposed for extending any single-neuron spike train kernel to multineuron spike trains, based on the R-convolution kernel. Special subclasses of the proposed R-convolution linear combination kernel are explored. These subclasses have a smaller number of parameters and make optimization tractable when the size of data is limited. The proposed kernel was evaluated using Gaussian process regression for multineuron spike trains recorded from an animal brain. It was compared with the sum kernel and the population Spikernel, which are existing ways of decoding multineuron spike trains using kernels. The results showed that the proposed approach performs better than these kernels and also other commonly used neural decoding methods.

A dictionary learning algorithm for sparse coding by the normalized bilateral projections

Sparse coding is a method of expressing the input vector as a linear combination of a few vectors taken from a set of template vectors, often called a dictionary or codebook. A good dictionary is the one that sparse codes most vectors in a given class of possible input vectors. There are currently several proposals to learn a good dictionary from a set of input vectors. Such methods are termed under the title of dictionary learning. We propose a new dictionary learning algorithm, called K-normalized bilateral projections (K-NBP), which is a modification to a widely used dictionary learning method, i.e., K-singular value decomposition (K-SVD). The main idea behind this was to standardize and normalize the input matrix as a preprocessing stage, and to correspondingly normalize the estimated source vectors in the dictionary update stage. The experimental results revealed that our method was fast, and when the number of iterations was limited, it outperformed K-SVD. Also, if only a coarse approximation was needed, it provided results that were almost like those from K-SVD, but with fewer iterations. This indicated that our method was particularly suited to large data sets with many dimensions, where each iteration took a long time.

Image Retrieval with Generative Model for Typicality

One of the most common image retrieval tasks is to find the most typical image that depicts the object specified by a query. Existing image search engines cannot efficiently do this since their search results are often a mixture of images belonging to various semantic concepts. We therefore introduce a probabilistic model for typicality. Our model consists of images, symbolic features, and latent semantic concepts (aspects). The aspect with highest probability is assumed to represent typicality. By collecting a large number of images, we can estimate parameters using EM algorithm. The estimated parameters are used to quantify the level of typicality for each image. Based on the proposed method, we have implemented a system, for ranking images by their typicality. Experiments using both artificial and real data showed the effectiveness of our method.

Web and Wireless Geographic Information Systems

W2GIS is a series of events continuously developing and expanding, with the aim to provide a forum for discussing advances in theoretical, technical, and practical issues in the field of wireless and Internet technologies suited for the spreading, using, and processing of geo-referenced data. Following a tradition of alternation between Europe and Asia, the 11th edition took place in the wonderful city of Naples, Italy, in April 2012, offering participants a unique setting for scientific discussion. The international program committee was composed of 38 representatives from 15 countries, and made a valuable effort to select 13 full papers and 4 short papers from a total of 32 submissions, coming from 12 countries in 4 continents. These papers have been published in a Volume of LNCS [1]. As a result, the technical program of the symposium provided many stimulating contributions, ranging from spatial humancomputer interaction to positioning, including sensor networks and geo semantics. Two invited speakers further enriched the event. Prof. Christopher Jones, from the School of Computer Science at Cardiff University (UK), presented an overview of a natural language photo captioning system able to describe the geographical context of the photo with regard to its possible subject, in association with proximal and regional place names. Dr. Andreas Sasse, from the Volkswagen Group Research, Department of Driver Information Systems, in Wolfsburg (DE), provided a stimulating talk on the future requirements, in terms of accuracy and updates, for maps and location-based services to be included in tomorrow’s vehicles. For this special feature, revised and extended versions were solicited from the highest quality papers at the 11th W2GIS symposium. The usual rigorous JOSIS review process was applied, leading to the acceptance of two papers in the field of sensor networks. Nowadays we live in an environment which is more and more embedded of sensors, able to capture and distribute observations of the phenomena surroundings our lives. Consequently, much research effort worldwide is expended on addressing the challenges posed by this new technological evolution. The first paper of this special feature, is by Annalisa Appice, Anna Ciampi, Donato Malerba, and Pietro Guccione. This article describes a new approach to deal with the problem of missing observations in networks of sensors. In particular, authors propose a novel spatiotemporal interpolation process, overcoming the tendency to treat space and time separately. The proposed technique is empirically evaluated using two large, real climate sensor networks. The results show that, in spite of a notable reduction in the volume of data, the technique guarantees accurate estimation of missing data. The second paper is by Mohamed Bakillah, Steve H.L. Liang, Alexander Zipf, and Mir

Visualization of N-Gram Input Patterns for Evaluating the Level of Divergent Thinking

Creative thinking is often considered to be difficult, one reason because humans tend to be trapped in the same patterns of thinking and cannot easily come up with a totally new combination of concepts. In other words, humans are not talented at evenly exploring combinatorial space. In order to visualize how strong this tendency is, we implemented a system that asks the subject to type in a long sequence of numbers. The system then counts the frequency of the appearance of the same subsequences using n-grams. We called it “the Creativity Test.” It measures one’s efficiency of exploring a wider part of a combinatorial space without being caught in few patterns. The result is assumed to be related to the ability of divergent thinking, which is considered to be important in creative thinking. When we tested the system on a group of subjects, we discovered that, for most of them, surprisingly long n-grams appeared frequently, making the subjects realize how inefficient they were at coming up with new combinations.