Ryohei Orihara

Activity Recognition for Dogs Based on Time-series Data Analysis

Dogs are one of the most popular pets in the world, and more than 10 million dogs are bred annually in Japan now [4]. Recently, primitive commercial services have been started that record dogs’ activities and report them to their owners. Although it is expected that an owner would like to know the dog’s activity in greater detail, a method proposed in a previous study has failed to recognize some of the key actions. The demand for their identification is highlighted in responses to our questionnaire. In this paper, we show a method to recognize the actions of the dog by attaching only one off-the-shelf acceleration sensor to the neck of the dog. We apply DTW-D which is the state-of-the-art time series data search technique for activity recognition. Application of DTW-D to activity recognition of an animal is unprecedented according to our knowledge, and thus is the main contribution of this study. As a result, we were able to recognize eleven different activities with 75.1 % classification F-measure. We also evaluate the method taking account of real-world use cases.

Japanese Text Classification by Character-level Deep ConvNets and Transfer Learning.

Temporal (one-dimensional) Convolutional Neural Network (Temporal CNN, ConvNet) is an emergent technology for text understanding. The input for the ConvNets could be either a sequence of words or a sequence of characters. In the latter case there are no needs for natural language processing that depends on a language such as morphological analysis. Past studies showed that the character-level ConvNets worked well for news category classification and sentiment analysis / classification tasks in English and romanized Chinese text corpus. In this article we apply the character-level ConvNets to Japanese text understanding. We also attempt to reuse meaningful representations that are learned in the ConvNets from a large-scale dataset in the form of transfer learning, inspired by its success in the field of image recognition. As for the application to the news category classification and the sentiment analysis and classification tasks in Japanese text corpus, the ConvNets outperformed N-gram-based classifiers. In addition, our ConvNets transfer learning frameworks worked well for a task which is similar to one used for pre-training.

Factors Affecting Accuracy in Image Translation based on Generative Adversarial Network.

With the development of deep learning, image translation has made it possible to output more realistic and highly accurate images. Especially, with the advent of Generative Adversarial Network (GAN), it became possible to perform general purpose learning in various image translation tasks such as “drawings to paintings”, “male to female” and “day to night”. In recent works, several models have been proposed that can do unsupervised learning which does not require an explicit pair of source domain image and target domain image, which is conventionally required for image translation. Two models called “CycleGAN” and “DiscoGAN” have appeared as state-of-the-art models in unsupervised learning-based image translation and succeeded in creating more realistic and highly accurate images. These models share the same network architecture, although there are differences in detailed parameter settings and learning algorithms. (in this paper we will collectively refer to them as “learning techniques”) Both models can do similar translation tasks, but it turned out that there is a large difference in translation accuracy between particular image domains. In this study, we analyzed differences in learning techniques of these models and investigated which learning techniques affect translation accuracy. As a result, it was found that the difference in the size of the feature map, which is the input for the image creation, affects the accuracy.

Do Professional Football Players Follow the Optimal Strategies in Penalty Shootout

Do people and companies choose optimal strategies under various situations? If not so, why not? Pursuing the reason for this helps to understand individuals and companies. In addition, game theory has been heavily involved in the understanding of sports, economics and other social sciences. In this study, we focused on football’s penalty shootout where data is relatively easy to collect, mixed strategy can be applied and making a pay-off matrix considering our own probability is possible. Using the pay-off matrix, we obtained optimal strategy of the kicker in the penalty shootout and revealed the gap between the optimal strategy and actual action taken by players. We compared the probability distribution for each data attribute in the dataset in order to obtain the cause of the gap. We use 100 professional penalty shootouts (total 1032 kicks) which were collected from internet video site during the period from 2001-2016. Experimental results showed that there was a gap between the optimal strategy and the actual action taken by players and that it also suggested the position and team attributes and temporary scores of the shootout and kicking order involved in the gap. Considering them we made the hypothesis and estimated the cause of the gap. We hope this method can apply to other fields than sports.

Text Classification and Transfer Learning Based on Character-Level Deep Convolutional Neural Networks

Temporal (one-dimensional) Convolutional Neural Network (Temporal CNN, ConvNet) is an emergent technology for text understanding. The input for the ConvNets could be either a sequence of words or a sequence of characters. In the latter case there are no needs for natural language processing. Past studies showed that the character-level ConvNets worked well for text classification in English and romanized Chinese corpus. In this article we apply the character-level ConvNets to Japanese corpus. We confirmed that meaningful representations are extracted by the ConvNets in English corpus and Japanese corpus. We attempt to reuse the meaningful representations that are learned in the ConvNets from a large-scale dataset in the form of transfer learning. As for the application to the news categorization and the sentiment analysis tasks in Japanese corpus, the ConvNets outperformed N-gram-based classifiers. In addition, our ConvNets transfer learning frameworks worked well for a task which is similar to one used for pre-training.

Fast Many-to-One Voice Conversion using Autoencoders.

Most of voice conversion (VC) methods were dealing with a one-to-one VC issue and there were few studies that tackled many-to-one / many-to-many cases. It is difficult to prepare the training data for an application with the methods because they require a lot of parallel data. Furthermore, the length of time required to convert a speech by Deep Neural Network (DNN) gets longer than pre-DNN methods because the DNN-based methods use complicated networks. In this study, we propose a VC method using autoencoders in order to reduce the amount of the training data and to shorten the converting time. In the method, higher-order features are extracted from acoustic features of source speakers by an autoencoder trained with source speakers’ data. Then they are converted to higher-order features of a target speaker by DNN. The converted higher-order features are restored to the acoustic features by an autoencoder trained with data drawn from the target speaker. In the evaluation experiment, the proposed method outperforms the conventional VC methods that use Gaussian Mixture Models (GMM) and DNNs in both one-to-one conversion and many-to-one conversion with a small training set in terms of the conversion accuracy and the converting time.

Sarcasm Detection Method to Improve Review Analysis.

Currently, classifying sarcastic sentences into positive and negative sentiments has been a difficult problem and an important task. The sarcastic sentences could indicate negative meaning by using positive expressions, or positive meaning by using negative expressions.Sarcasm is a special kind of sentiment that comprise of words which mean the opposite of what you really want to say, especially in order to insult or wit someone, to show irritation, or to be funny.Therefore, determining sarcasm is an important task in order to correctly classify the sentence. In this paper, we propose an approach to detect sarcasm.First, we apply dependency parsing to amazon review data. After that, we classify phrases in the sentence into the proposed phrase based on the sequence of part-of-speech as proposed by Bharti et al. After being classified into either one of the phrase types, it is determined whether each phrase is positive or negative.If the emotions of the situation phrases and the sentiment phrases are different, the sentence is determined to be a “sarcasm”. Using the above method , the experimental result shows the effectiveness of our method as compared with the the existing research.

Activity Recognition for Dogs Using Off-the-Shelf Accelerometer

Dogs are one of the most popular pets in the world, and more than 10 million dogs are bred annually in Japan now (JPFA, 2013). Recently, primitive commercial services have been started that record dogsâ?? activities and report them to their owners. Although it is expected that an owner would like to know the dogâ??s activity in greater detail, a method proposed in a previous study has failed to recognize some of the key actions. The demand for their identification is highlighted in responses to our questionnaire. In this paper, we show a method to recognize the actions of the dog by attaching only one off-the-shelf acceleration sensor to the neck of the dog. We apply DTW-D which is the state-of-the-art time series data search technique for activity recognition. Application of DTW-D to activity recognition of an animal is unprecedented according to our knowledge, and thus is the main contribution of this study. As a result, we were able to recognize ten different activities with 65.8% classification F-measure.

Towards the Elimination of the Miscommunication Between Users in Twitter

In recent years, a Twitter response from another user who does not share the intentions and expectations of the original poster may cause discomfort and stress, which is a social phenomenon known as SNS fatigue. For example, a user may receive answers that are different from her/his expectation after the user posts a question on the timeline. In the background of such responses there is a miscommunication between users. In order to resolve the problem, it is important to know what the original poster expected as responses to her/his tweet. In this paper, we propose a classification method of tweets according to the response that users expect, and experimentally evaluate it. As a result, we have shown that tweets which the poster does not expect any replies can be classified with 76.2 % of the average precision.

Decision Making Strategy Based on Time Series Data of Voting Behavior

In gambling such as horse racing, we are sometimes able to peep peculiar voting behavior by a punter with the advantageous information closely related to the results. The punter is often referred as an insider. In this study, our goal is to propose a reasonable investment strategy by peeping insiders’ decision-making based on the time series odds data in horse racing events held by JRA. We have found the conditions that the rate of return is more than 642 % for races whose winner’s prize money is 20 million yens or more. That suggests the possibility of Knowledge Peeping.