The Transformer is a widely used neural network architecture for natural language processing. Recently, it has been applied to time series prediction tasks. However, the vanilla transformer has a critical limitation in that it cannot predict the time intervals between elements. To overcome this limitation, we propose a new model architecture called TI-former (Time Interval Transformers) that predicts both the sequence elements and the time intervals between them. To incorporate the elements’ sequential order and temporal interval information, first we propose a new positional encoding method. Second, we modify the output layer to predict both the next sequence element and the time interval simultaneously. Lastly, we suggest a new loss function for timestamped sequences, namely Time soft-DTW, which measures similarity between sequences considering timestamps. We present experimental results based on synthetic sequence data. The experimental results show that our proposed model outperforms than vanilla transformer model in various sequence lengths, sequence numbers, and element occurrence time ranges.

Graph Neural Networks (GNNs) are powerful tools for analyzing graph-structured data in various fields because of their great expressive power for graph data. They use a message-passing mechanism to update node embeddings, which are then used for tasks such as node classification and link prediction. Recently, node embeddings have also been used in research on graph node clustering, which aims to group similar nodes based on their features and graph topology. However, traditional methods for node clustering have a limitation in that GNNs only focus on generating node embeddings without considering the ultimate objective of clustering. To address this issue, a novel technique called "Deep Clustering" has been proposed, which integrates both node embedding and clustering stages. This requires defining a new loss function by simultaneously minimizing the GNN loss and the clustering loss. Our proposed loss function incorporates not only the distance within clusters but also the distance between clusters by applying the Silhouette coefficient, which enables us to achieve better clustering results. In this paper, we propose a Silhouette-based Deep Clustering with GNN (SDCG) to more effectively cluster nodes in a graph by iteratively training the embedding model to produce embedding vectors with improved clustering results. Through extensive experiments, we demonstrate that SDCG outperforms the conventional approach of performing embedding and clustering independently.

The following topics are dealt with: learning (artificial intelligence); security of data; mobile computing; Internet; computer aided instruction; feature extraction; health care; Internet of Things; neural nets; and cloud computing.

Biometrics are widely used for security authentication systems to verify a person's identity such as fingerprint, iris, face, and voice recognition. Among them, unlike other biometrics, human gait has the advantage that it can be captured in an unobtrusive manner. In our previous research, we proposed a method of modeling the body parts of a captured walking person using the Kinect depth cameras. In this paper, we propose a new human gait recognition method that uses gait features extracted from the modeled body parts to identify a walking person. The proposed method uses a combination of static and dynamic gait features to improve the accuracy of person identification. Because each gait has a different cycle length, we also use a time normalization technique to transform gait feature sequences with different lengths to those of the same length to compare them more precisely. Based on the time-normalized gait feature sequences, we build a k-NN classifier and an LSTM classifier to classify different walking persons. Our experimental results show the high potentiality of the proposed method for identifying unknown walking persons.

Materialized views are nowadays commonly used in the data warehouse environment. Materialized views need to be updated when data sources change. Since the update of the views may impose a significant overhead, it is essential to update the views efficiently. Though there has been much work on efficient maintenance of a single view, maintenance of multiple views has not been sufficiently investigated.In this paper we propose an efficient incremental maintenance of multiple join views. In our previous work[6], we proposed the delta propagation strategy that computes the change of a join view in a recursive manner. We extend the delta propagation strategy to multiple views. The recursive property of the strategy makes it possible to share common intermediate results among views effectively. We first define the multiple view maintenance problem, then a heuristic algorithm that finds a global maintenance plan for the given views is proposed. We also present experimental result that shows the efficiency of the proposed method.

In this paper, a new nonlinear prediction model based on the Recurrent Neural Predictive Hidden Markov Model (RNPHMM) is proposed for speech enhancement. Assuming that speech is an output of the RNPHMM combining RNN and HMM, the proposed nonlinear prediction model-based recurrent neural network (RNN) is used to present the nonlinear and nonstationary nature of speech. The RNPHMM is a nonlinear prediction process whose time-varying parameters are controlled by a hidden Markov chain. Given some speech data for training, the parameters of the RNPHMM are estimated by a learning algorithm based on the combination of Baum-Welch algorithm and RNN learning algorithm using the back-propagation algorithm. In our experiment, the proposed method achieved about 2–2.5 dB of improvement in SNR compared with both the NPHMM and the HFM at various input SNRs.

In the data warehouse environment, the concept of a materialized view is nowadays common and important in an objective of efficiently supporting OLAP query processing. Materialized views are generally derived from select-project-join of several base relations. These materialized views need to be updated when the base relations change. Since the propagation of updates to the views may impose a significant overhead, it is very important to update the warehouse views efficiently. Though various view maintenance strategies have been discussed so far, they typically require too much access to base relations, resulting in the performance degradation.In this paper we propose an efficient incremental view maintenance strategy called delta propagation that can minimize the total size of base relations accessed by analyzing the properties of base relations. We first define the delta expression and a delta propagation tree which are core concepts of the strategy. Then, a dynamic programming algorithm that can find the optimal delta expression are proposed. We also present various experimental results that show the usefulness and efficiency of the strategy.

In this paper, a gain-adapted speech recognition method in unknown noise is developed in the time domain. Noise is assumed to be colored. To cope with the notable nonstationary nature of speech signals such as fricative, glides, liquids, and transition region between phones, the nonstationary autoregressive (NAR) hidden Markov model (HMM) is used for clean speech. The nonstationary AR process is represented by using polynomial functions with a linear combination of M known basis functions. When only noisy signals are available, the estimation problem of unknown noise inevitably arises. By using multiple Kalman filters, the estimation of noise model and gain contour of speech is performed.

A mixture interacting multiple model (MIMM) algorithm is proposed to enhance speech contaminated by additive nonstationary noise. In this approach, a mixture hidden filter model (HFM) is used for clean speech modeling and a single hidden filter is used for noise process modeling. The MIMM algorithm gives better enhancement results than the IMM algorithm. The results show that the proposed method offers performance gain compared to the previous results in with slightly increased complexity.

A time-domain approach for enhancing speech signals degraded by statistically independent additive nonstationary noise with no a priori information is developed. The autoregressive (AR)-hidden filter model (HFM) with gain contour is proposed for modeling the statistical characteristics of the clean speech signal. Given the HFM parameter set of the speech, speech enhancement becomes a set of problems of joint signal estimation for clean speech and system identification for the gain contour and time-varying parameter of noise. Then, the expectation-maximization (EM) algorithm is applied to signal estimation and system identification. In the E-step, the signal estimation becomes a weighted sum of conditional mean estimator using multiple Kalman filters with Markovian switching coefficient, where the weights equal to a posteriori probabilities of the specific state sequence history given the noisy speech. The probability is computed by the Viterbi algorithm (VA). In M-step, the gain contour and noise parameters are recursively updated by an adaptive algorithm modified from the gradient-based algorithm. The proposed method does not require framing of speech signal in, the train and enhancement procedure. The proposed method is tested against the noisy speech signals degraded by nonstationary noise at various input signal-to-noise ratios. An approximate improvement of 4.5-6.0 dB in signal-to-noise ratio (SNR) is achieved at the input SNR 10 and 15 dB.

A gain-adapted speech recognition in unknown noise is developed in time domain. The noise is assumed to be the colored noise. The nonstationary autoregressive (NAR) hidden Markov model (HMM) used to model clean speeches. The nonstationary AR is modeled by polynomial functions with a linear combination of M known basis functions. Enhancement using multiple Kalman filters is performed for the gain contour of speech and estimation of noise model when only the noisy signal is available.

We consider an adaptive filtering algorithm for speech parameter estimation and enhancement when the observation noise is colored with no a priori information. The resulting algorithm consists of adaptive filtering procedures that recursively estimate and enhance the parameters of speech plus noise model, and results in /spl sim/3 dB improvement over Gaussian assumption on the excitation source.

This paper considers speech enhancement where the speech signal is modeled with hidden filter models (HFMs), when only noisy speech signal are available. The HFM is a parametric approach for representing the speech waveform in the time domain. We apply the nested EM algorithm for jointly estimating the clean signal and the parameters of HFM and noise model. A computationally efficient implementation of the algorithm is developed by the log-likelihood gradient based on the Kalman filter output in the estimation process. The resulting algorithm does not need framing and may be viewed in the time domain context.

A recursive estimation to enhance speech additively contaminated by colored noise is proposed. This method is based on the Kalman filter with time-varying modeling of the clean speech signal. Then, a hidden filter model is used to model the clean speech signal. An improvement of approximately 4.3 dB and 3.2 dB in signal-to-noise ratio (SNR) is achieved at 10 dB and 15 dB input SNR, respectively.

In many applications we have to deal with densities which are highly non-Gaussian or which may have Gaussian shape in the middle but have potent deviations in the tails. To fight against this deviation, we consider Bayesian estimation in the case of non-Gaussian state and measurement noises. The robust estimation problem for linear discrete-time systems is considered here. The non-Gaussian noises are modeled as a mixture. We present a robust recursive estimation model that is an approximate minimum variance estimator with a maximum a posteriori (MAP) decision rule for determining the noise sequence distribution.>

There are two major problems in estimating vocal tract characteristics by conventional linear prediction: estimation accuracy being subject to the characteristics of the excitation source, and the output quality and the estimation accuracy deteriorating with additive background noise. The authors solved these problems as follows: first, estimate the parameters of a robust AR model where the driving noise is a mixture of a Gaussian and an outlier process; then, propose an iterative procedure that involves parameter estimation for uncorrupted speech and data cleaning based on the robust Kalman filter; lastly, the above results are used to enhance speech corrupted by white noise. The results are more efficient and less biased for uncorrupted speech, and superior at low SNR for noisy speech.>

In the multipulse linear predictive coding (LPC) speech synthesis, an autoregressive filter is excited by a multipulse excitation consisting of impulses of various amplitudes and locations. The authors statistically model the excitation signal as a zero mean Bernoulli-Gaussian process. The pulse locations are independently distributed with a probability distribution, and the pulse amplitudes are expressed as a Gaussian sequence with zero mean and finite variance. An algorithm is described for estimation of pulse amplitudes and locations based on the Bernoulli-Gaussian process when the number of pulses is given. Results from computer simulation are presented.>