Report on efforts to optimize the supply chain in fresh food as one of the use cases for solving social problems where the application of artificial intelligence technology is expected from AITC(Artificial intelligence technology consortium). We aim to produce a cyber market. First, we construct a consumption prediction model that utilizes weather data etc. Next, the demand amount plus the business condition is determined, and the growth prediction model is coordinated.

In this study, we tried to estimate child's interest as a preliminary investigation towards introduction of AI technology to childcare. The research method focused attention on the child's interest that the childcare professional would estimate in nursery practice.In addition, we conducted annotations by teachers on child's interests by sensing daily childcare scenes.As a result, the interest of 18 children was described. Then, it was suggested that the physical activity state of the child is involved in the estimation of the child's interest. From these results, usefulness of measuring physical activity was recognized for estimating children's interests. We also showed that automatic estimation of child's interest using AI technology could be realized through measurement of physical activity state.

There are many big-scale events all around to encourage learning sciences, recruiting or advertising. Though these events hosts want these events to be more meaningful and interesting to both visitors and exhibitors, it is difficult to correct data to improve the management of the events for their short period or rent space. We show how correct visitors' data easily and use these data and artificial technology to improve these events.

A method of data sampling from a huge data set is discussed. We introduce a generalized relative square error to emphasize low probability events and figure out the best sampling weight to reduce the error. Our arguments are based on the large deviation theory. Large reduction in the generalized relative square error was numerically confirmed for the best sampling weight. We also propose to use Wang-Landau algorithm in data sampling. This algorithm is not only efficient to estimate a distribution of the original data, but also useful in data sampling to suppress the statistical errors.

In this study, we aim to find effective and personalized questions from a questionnaire that enable answerers clustering. We propose a question selection method based on a Bayesian network model that is constructed from probabilistic clustering results. We show that the gap between the accuracies of estimating cluster index of our method and the upper bound of them is very small.

In this research, we propose a method incorporating structure of ordered logit model into output layer of Self-Attentive-LSTM. This helps to predict ordinal variables. Model with using this method learns relationships among classes since the order of classes are assumed in advance. We show this model is useful when the objective variable is ordinal by applying to "The economic watcher", data surveyed by the Japanese Cabinet Office. When predicting the Economic assessment index, this proposed method improved F-measures of both sides of the ordinal indices compared to Self-Attentive-LSTM.

Visualizing chains of Economic events and Financial matters including background eventsfrom text data is useful for investors and manager to understand Economic events and Financial matters properly. However, extracting chains of Economic events and Financial matters manually takes a lot of time. Therefore, we treat chains of Economic events and Financial matters as chains of causal relations (we call them Causal Network) and propose the procedure creating causal network using vectors similarity which represent semantic similarities between expressions of Economic events and Financial matters.

In recent years, time series data have been collected in many fields, and visual analytics interface is expected to be useful for utilizing such data. However, several issues arising when applying it to time series data should be considered. For example, when time series data is visualized using animation, collision would occur between movement of time series data itself and movement caused by interaction with users. In order to solve those issues, this paper proposes a visual analytics interface for time series data based on trajectory manipulation, which can handle temporal and spatial changes uniformly. The results of user experiment show the effectiveness of the proposed interface.

As the evolution of sensors and computers enables collecting abundant data, methods to analyze high-dimentional data are becoming important. Dyanmic mode decompostion (DMD) is a data-driven method to extract dynamic structure from data and is attracting attention recently. In this study, we made use of DMD to analyze sound data of rotary machines with normal and abnormal states. We applied DMD to spectral distributions of the data and analyzed the dynamic structure of spectral distributions. We found that on spectral distributions of data from abnormal states, time-decaying structure is more likely to be dominant than those from normal states.

Companies develop various activities to manage brands as assets in order to acquire target brand images. However, in fact, since companies cannot grasp the necessary elements for the image, there are companies that make inconsistent products and promotions into the world. Therefore, in this research, we verified factors to form "quality" brand image, which is said to be ambiguous and complicated, in customers’ perception. Quality is said to include not only objective value (Functional value) such as performance and durability but also subjective value (Emotional value) such as beauty and perceived quality. In recent years, companies such as Apple and Samsung, which have excellent emotional value, are emerging, so the value is considered as a major source of competition in the manufacturing industry. We believe that this research will enable companies to effective decision-making without obscuring the elements necessary to acquire the target brand image.

With the advent of social media, social bots—computer programs that automatically post news—have become popular, and some of these have been purposefully used to disseminate fake news. This is an increasingly serious problem in the digital era. Although a lot of bot detection methods have been proposed, such methods often cannot overcome their language dependencies and fail to find bot clusters. This research aims at constructing a language-independent method for detecting social bots and their clusters on Twitter using machine learning. The results provide new information available for the development of such methods.

Social media is not only a sensor but also an actuator. A social sensor can detect various social phenomena or trends by using the data obtained from social media. Though the information from social sensors has vastly enhanced our potential to observe and predict real-world phenomena, the causality from the social media itself to the real world has not been studied to date. Meanwhile, recent trends in SNS-driven consumer behavior, such as taking beautiful photos for Instagram, so-called "Instagenic," have further highlighted the importance of studying the causality from social media to the real world. This paper demonstrates a new concept of social actuator. We introduce internal states, which represent the states of the social media users influenced by others, and show how to address the confounding structure in the inference of causality from social media to the real world. Using the results of our experiments on Twitter data and cryptocurrency-market data, we show that our proposed method can detect the influences between the users on social media, and describe the causation from Twitter to the cryptocurrency market. Finally, we discuss the effectiveness of the proposed method for different datasets and suggest that we all have the potential to impact the real world through social media either intentionally or unintentionally.

We propose to find representative data points from continuous data via a two-step procedure: We first binarize data points based on the nearest neighbor search, followed by performing frequent pattern mining on the binarized data. Since frequent patterns correspond to combinations of data points shared by many other data points as their neighbors, they are expected to well summarize the entire dataset. We empirically show that representative data points detected by our method have competitive quality with random sampling in the classification scenario.

Boiling point is one of the most basic physical properties of chemical compounds. If it is possible to know the boiling point, it can be used to identify an unknown compound. The boiling point can be also used to predict other physical properties of the compound. Thus, estimation of boiling point is one of the important research areas in computational chemistry, and is still being studied actively. It is empirically known that there is a close relationship between chemical structure and physical properties, so that physical properties can be thought of as a function of chemical structure. On other hand, various graph invariants derived from molecular graphs can also be considered as a function of chemical structure. In this study, we tried to construct a prediction model by multiple linear regression analysis using eigenvalues obtained from matrix representations of molecular graphs. For 51 chemical compounds of saturated hydrocarbons, a regression model obtained with the first four largest eigenvalues and the number of carbon atoms gave us good estimates of their boiling points.

The acquisition of the CAD operational skills is necessary to effective car development. In this study, we tried extraction of the operational skills by analysis of CAD operation log. As a result, we succeeded in extracting the tacit knowledge from CAD operation log.

We present a stochastic gradient descent for the time-series forecasting problem in the presence of concept drift.

Sparsity-promoting dynamic mode decomposition (SP-DMD) is a data-driven method for estimating a modal representation of a nonlinear dynamical system, where the modes are selected via l1-regularization depending on the tradeoff between the quality of the representation and the number of the modes. However, the way to statistically evaluate modes selected by SP-DMD is not established. If statistical evaluation is not performed, we may not specify issues caused by different reasons such as noise and overfitting. In this paper, we propose a method to statistically evaluate modes selected by SP-DMD. We develop the method based on the combination of bootstrap and SP-DMD.

The training methods of sequence generator with a combination of GAN and policy gradient has shown good performance. In this paper, we propose expert-based reward function training: the novel method to train sequence generator. Different from previous studies of sequence generation, expert-based reward function training does not utilize GAN's framework. Still, our model outperforms SeqGAN and a strong baseline, RankGAN.

Dynamic mode decomposition(DMD) is a data-driven method for representing high-dimensional, nonlinear dynamical systems. DMD extracts key low-rank spatiotemporal features of the high-dimensional systems. However, since DMD is an unsupervised method and, thus, cannot incorporate label information into it even when such information is available. In this paper, we propose a framework to incorporate supervised information into DMD analyses. Experimental results show the effectiveness of performing classication tasks using modes obtained by the proposed method.

Neural networks have rich ability to learn complex representations. However, due to the limited number of training samples, overfitting is likely to occur. Hence, it is essential to regularize the learning process of neural networks. In this paper, we propose a regularization method which estimates CNN's parameters as probabilistic distributions by using hypernet. Then, to make it applicable to a large model such as WideResNet, we used likelihood as loss function. Experimental results demonstrate the regularization of our method.

This paper aims to verify robustness of covariance matrix adaptation evolution strategy (CMA-ES) optimization for neural networks (NN). We added label noise to the training dataset. Unlike stochastic gradient descent (SGD), which is the state-of-the-art optimizer of NN, a CMA-ES based optimization was robust against label noise.

SUNA is currently one of the most adaptive neuroevolution methods which is able to tackle different problems efficiently. However, many questions remain unanswered. In this research, we applied SUNA to the bipedal-walking problem and evaluate it general learning properties. The results show that even without any modificiations SUNA is able to learn in this environment. Moreover, contrary to many other methods, it is continuously improving its average rewards showing a near open-ended learning.

Probabilistic topic models based on latent Dirichlet Allocation is widely used to extract latent topics from document collections. In recent years, a number of extended topic models have been proposed, especially Gaussain LDA(G-LDA) has attracted a lot of attention. G-LDA integrates topic modeling with word embeddings by replacing discrete topic distribution over word types with multivariate Gaussian distribution on the word embedding space. This can reflect semantic information into topics. In this paper, we use a G-LDA for our base topic model and apply Stochastic Variational Inference (SVI), an efficient inference algorithm, to estimate topics. This encourages the model to analyze massive document collections, including those arriving in a stream.

In both supervised and unsupervised learning tasks, embedding the underlying data into a function space using a "kernel mean" has been well-studied, and is known to be an efficient means of characterizing even complex distributions. Here we consider a broad generalization of this notion to countless "functional parameters" of the underlying distribution, and as a concrete example explore what may naturally be called the "kernel median" of the data. In this short paper, we formulate the new parameter class, provide a procedure for obtaining an important special case, with basic convergence guarantees and expressions useful for practical implementation.

This paper discusses the way to sort items by the noisy feedback of pairwise comparisons. An algorithm that partially sorts items are provided, and its empirical performance is shown.

We propose a new loss function based on variance of generated images that we introduce to GAN(Generative Adversarial Networks). Recent image-generation methods adopt a neural-network based generative model. Auto- generators of illustrations can contribute to assist creative activities and entertainment. This paper focuses on supression of collapse and its benefit to GAN training. Using our new technique, we attempted to generate various images of illustrations.

In a general decision-making task, the options of action are expanded indefinitely due to the change of environment or the discovery of new action by agent. In a situation that the number of options increase, it is necessary for an agent to acquire an abstracted expression of actions autonomously. Here we propose a learning framework that solve this issue. In the proposed method, value function is approximated with embedded behavioral representations, which generalize the expression of actions, using state-tradition trajectories. We confirmed the efficiency of the framework using the mobile game "Gyakuten Othellonia". This game is a mixture of board game and trading card game and characters are added to the environment frequently, which is a good testbed to realize expandable action space. Finally, we show that, with the proposed framework, an agent can learn character's representation and utilize it to learn optimal strategies in the game.

We present a time convolutional variational ladder autoencoder (TCVLAE), which learns disentangled and interpretable representations from sequential data without supervision. For the simple 2d data set, the proposed model experimentally shows that it is possible to separate the meaning of time series data.

As deep layered neural networks enables reinforcement learning in huge action-state spaces, the exploration--exploitation tradeoff becomes more serious. Several heuristics have been proposed to deal with the tradeoff utilizing noises. The probabilistic methods have difficulty in parameter tuning, and they amplify the problem of huge dispersion in performance of deep reinforcement learning algorithms. We propose a deterministic action selection algorithm based on a cognitive satisficing value function (RS) inspired by how humans explore under uncertainty. We define a method to enable optimal (minimal) exploration, utilizing the relationship between the aspiration level and the potential exploration distribution. The resulting algorithm exhibits an optimal performance in multi-armed bandit problems, and it opens the possibility for a new class of reinforcement learning algorithms.

In recent years, a lot of analytical methods of medical images using deep learning are suggested. Especially, convolutional neural network (CNN) is a model generally used in image recognition. When we classify diffuse lung disease (DLD) patterns using CNN, it is necessary to set region-of-interests (ROIs) on CT images. However, detection is important on diagnosis of DLD as same as classification. So, we propose a method to detect DLD opacities and extract DLD areas without setting ROIs. In this study, we evaluated detection methods of DLD areas using CNN, FCN and U-Net.

Volcanic eruptions sometimes cause severe damage to many people. This paper explains our attempts for predicting volcanic eruptions from time series sensor data obtained from volcanic monitoring systems (strainmeters) located in Sakurajima. Given the time series data of strainmeters for 100 minutes, our goal is to predict future status of the volcano which is either "explosive" or "not explosive" for the 60 minutes immediately after the 100 minutes. We use stacked recurrent neural network for this task, and our method achieves 66.1% F-score on average. We also propose a four-stage warning system that classifies time series sensor data into the following categories: "Non-eruption", "May-eruption", "Warning" and "Critial". The percentage of "explosive" cases in "Critial" category is 51.9%.

One of the most common needs in manufacturing plants is rejecting products not coincident with the standards as anomalies. Manufacturing companies usually employ numerous inspectors for anomaly detection and it takes a high cost. Accurate and automatic anomaly detection reduces inspection cost and improves product reliability. In unsupervised anomaly detection, a probabilistic model detects test samples with lower likelihoods as anomalies. Recently, a probabilistic model called deep generative model (DGM) has been proposed for end-to-end modeling of natural images and already achieved a certain success. However, anomaly detection of machine components with complicated structures is still challenging because they produce a wide variety of the normal image patches with lower likelihoods. For overcoming this difficulty, we propose unregularized score for the DGM. As its name implies, the unregularized score is the anomaly score of the DGM without the regularization terms. The unregularized score is robust to the inherent complexity of a sample and has a smaller risk of rejecting a sample appearing less frequently but being coincident with the standards. Experimental results of anomaly detection on the real manufacturing datasets show better performance of the unregularized score compared to existing approaches.

If some classes of the data have an only small number of samples, the accuracies of the classes become too low. It is well known as an imbalanced data problem. We often encounter imbalanced data in reality. In a sense, all the wild data are imbalanced. In this paper, we make pseudo-feature from feature map in lower layers of deep neural networks, and we augment the data of minor classes to improve the imbalanced-data problem. We compare our proposed method with existing ones in imbalanced data multi-class image classification problems.

Reducing garbage generated in our daily life is a key aspect to fulfill a sustainable human society. The lack of an effective method to obtaining information regarding the collected garbage, such as where, when and what kinds of waste are collected, is one of the major obstacles to promoting the garbage reduction in cities. In this research, we propose a system to automatically count the number of garbage bags, by using Convolutional Neural Networks (CNN)-based technology to process the videos of garbage collecting process taken by a camera mounted at the back of garbage trucks. The proposed system is evaluated using the realistic garbage collecting videos of Fujisawa city. The experiment result shows that the proposed method achieves a high recall score of 0.90.

This paper proposes a novel method for training neural networks with a limited amount of training data. Our approach is based on knowledge distillation that transfers knowledge from a deep reference neural network to a shallow target one. The proposed method employs this idea to mimic predictions of non-neural networks reference models that are more robust against overfitting that the target neural network. Different from almost all the previous work for knowledge distillation that requires a large amount of labeled training data, the proposed method requires only a small amount of training data. Instead, we introduce pseudo training data that is optimized as a part of model parameters.

Contract farming has a managerial advantage that farmers can directly negotiate prices with business partners. And it is necessary to predict harvesting date of agricultural crops precisely for contract farming. Conventionally, this precondition has been solved only by experience rule of experienced farmers. However, for new farmers who do not have experiential rules or for large-scale field managers, contract farming is difficult management method. In order to solve this problem, we proposed automatic harvesting date prediction method using statistical model by deep learning. We confirmed that deep learning models exceed the accuracy of non-deep learning models.

In recent years, complex machine learning models like deep neural networks play a central role in many real applications, due to their high predictive performance. Interpreting machine learning models is then considered to be important since practitioners constantly need clues for improvement on such complex models, whose behavior is not directly visible to human. In this paper, we focus on the inner workings of convolutional neural networks, visualize them by a method called layer-wise relevance propagation, and report several findings from the visualization.

In this paper we extend the arithmetic operations of the Neural Programmer-Interpreters (NPI). NPI is a recurrent and compositional neural network that learns to represent and execute programs. First, we enable NPI to execute not only the addition that NPI was originally possible but also the other three arithmetic operations, i.e. subtraction, multiplication and division. Then, we extended NPI to make it possible to share subprograms between tasks for improving learning efficiency. Next, we solve word algebra problems for elementary school-level mathematics which can be solved by using four arithmetic operations. For this purpose, we develop a converter that converts word algebra problems into mathematical expressions. This neural network is based on the Sequence-to-Sequence model with the attention mechanism. Using this neural network and NPI, we solve the data sets of word algebra problems and show that the accuracy of our method is better than the other existing methods.

We propose a novel deep learning framework to bidirectionally translate between robot actions and their linguistic descriptions. The model consists of two recurrent autoencoders, one of which is employed to make vector representations of robot actions and the other is for descriptions. The learning algorithm produces representations shared between actions and their descriptions by creating an additional loss function in which the representation of an action and that of its description become close to each other in the latent vector space. Across the shared representation, the trained model can produce a linguistic description given a robot action. The model is also able to generate an appropriate action by receiving a linguistic instruction, conditioned on the current visual input. Visualization of the shared representations shows that the robot actions are represented in a semantically compositional way in the vector space by being learned jointly with their descriptions.

In recent years, there are many researches of deep reinforcement learning to realize autonomous motion of robots. In deep reinforcement learning, a large number of trials such as thousands of times or more are required to realize sufficient performance as a learning result. However, learning in a real environment often requires assistance by people, so it is difficult to do thousands of trials. In this research, we create a learning database from efficient reinforcement learning that utilizes knowledge about tasks given by people in advance, and realize learning with a relatively small number of trials by performing mini batch learning using that database. We apply our proposed method to learning of picking task in the logistics warehouse and show the usefulness of our proposed method by comparing the results with other methods.

In recent years, the necessity for robots working in society has been spreading as the aging society has come. To easily be able to communicate with robots, it is expected that they can understand natural language and learn how to behave spontaneously through the interaction with humans. In this study, we aim to ground the meaning of natural language onto their behaviors by using reinforcement learning. In particular, we have proposed an efficient method to learn robot's motion with deep reinforcement learning by descritizing a robot's motion into a hierarchical structure consisting of basic motion elements which can be represented by words.