Recent efforts in learned cardinality estimation (CE) have substantially improved estimation accuracy and query plans inside query optimizers. However, achieving decent efficiency, scalability, and the support of a wide range of queries at the same time, has remained questionable. Rather than falling back to traditional approaches to trade off one criterion with another, we present a new learned approach that achieves all these. Our method, called ASM, harmonizes autoregressive models for per-table statistics estimation, sampling for merging these statistics for join queries, and multi-dimensional statistics merging that extends the sampling for estimating thousands of sub-queries, without assuming independence between join keys. Extensive experiments show that ASM significantly improves query plans under a similar or smaller overhead than the previous learned methods and supports a wider range of queries.

We propose a multi-singer emotional singing voice synthesizer, Muse-SVS, that expresses emotion at various intensity levels by controlling subtle changes in pitch, energy, and phoneme duration while accurately following the score. To control multiple style attributes while avoiding loss of fidelity and expressiveness due to interference between the attributes, Muse-SVS represents all attributes and their relations together by a joint embedding in a unified embedding space. Muse-SVS can express emotional intensity levels not included in the training data through embedding interpolation and extrapolation. We also propose a statistical pitch predictor to express pitch variation according to emotional intensity, and a context-aware residual duration predictor to prevent the accumulation of phoneme-level duration prediction errors, which is crucial for synchronization with instrumental parts. In addition, we propose a novel ASPP-Transformer, which combines atrous spatial pyramid pooling (ASPP) and Transformer, to improve fidelity and expressiveness by referring to broad contexts. In experiments, Muse-SVS exhibited improved fidelity, expressiveness, and synchronization performance compared with baseline models. The visualization results show that Muse-SVS effectively express the variation in pitch, energy, and phoneme duration according to emotional intensity. To the best of our knowledge, Muse-SVS is the first neural SVS capable of controlling emotional intensity.

We propose an end-to-end speech synthesizer, Fast DCTTS, that synthesizes speech in real time on a single CPU thread. The proposed model is composed of a carefully-tuned lightweight network designed by applying multiple network reduction and fidelity improvement techniques. In addition, we propose a novel group highway activation that can compromise between computational efficiency and the regularization effect of the gating mechanism. As well, we introduce a new metric called elastic mel-cepstral distortion (EMCD) to measure the fidelity of the output mel-spectrogram. In experiments, we analyze the effect of the acceleration techniques on speed and speech quality. Compared with the baseline model, the proposed model exhibits improved MOS from 2.62 to 2.74 with only 1.76% computation and 2.75% parameters. The speed on a single CPU thread was improved by 7.45 times, which is fast enough to produce mel-spectrogram in real time without GPU.

To address the issue of catastrophic forgetting in neural networks, we propose a novel, simple, and effective solution called neuron-level plasticity control (NPC). While learning a new task, the proposed method preserves the existing knowledge from the previous tasks by controlling the plasticity of the network at the neuron level. NPC estimates the importance value of each neuron and consolidates important neurons by applying lower learning rates, rather than restricting individual connection weights to stay close to the values optimized for the previous tasks. The experimental results on the several datasets show that neuron-level consolidation is substantially more effective compared to connection-level consolidation approaches.

We present a novel high-fidelity real-time neural vocoder called VocGAN.A recently developed GAN-based vocoder, MelGAN, produces speech waveforms in real-time.However, it often produces a waveform that is insufficient in quality or inconsistent with acoustic characteristics of the input mel spectrogram.VocGAN is nearly as fast as MelGAN, but it significantly improves the quality and consistency of the output waveform.VocGAN applies a multi-scale waveform generator and a hierarchically-nested discriminator to learn multiple levels of acoustic properties in a balanced way.It also applies the joint conditional and unconditional objective, which has shown successful results in high-resolution image synthesis.In experiments, VocGAN synthesizes speech waveforms 416.7x faster on a GTX 1080Ti GPU and 3.24x faster on a CPU than realtime.Compared with MelGAN, it also exhibits significantly improved quality in multiple evaluation metrics including mean opinion score (MOS) with minimal additional overhead.Additionally, compared with Parallel WaveGAN, another recently developed high-fidelity vocoder, VocGAN is 6.98x faster on a CPU and exhibits higher MOS.

In capsule networks, the routing algorithm connects capsules in consecutive layers, enabling the upper-level capsules to learn higher-level concepts by combining the concepts of the lower-level capsules. Capsule networks are known to have a few advantages over conventional neural networks, including robustness to 3D viewpoint changes and generalization capability. However, some studies have reported negative experimental results. Nevertheless, the reason for this phenomenon has not been analyzed yet. We empirically analyzed the effect of five different routing algorithms. The experimental results show that the routing algorithms do not behave as expected and often produce results that are worse than simple baseline algorithms that assign the connection strengths uniformly or randomly. We also show that, in most cases, the routing algorithms do not change the classification result but polarize the link strengths, and the polarization can be extreme when they continue to repeat without stopping. In order to realize the true potential of the capsule network, it is essential to develop an improved routing algorithm.

Deep1 Convolutional Neural Network (DCNN) is a break-through technology in image recognition. However, because of extreme computing resource requirements, DCNN need to be implemented by hardware accelerator. In this paper, we present an FPGA-based accelerator design techniques of DCNN for handwritten Hangul character recognition engine. We achieved about 11.9ms recognition time per character with Xilinx FPGA accelerator. Our design optimization was performed with Xilinx HLS and SDAccel environment targeting Kintex XCKU115 FPGA from Xilinx. Our design outperforms CPU in terms of execution time 6.25 times, and GPGPU in terms of energy efficiency 4.7 times and cooling cost for the computing servers by 17 times. We think the research results imply deep learning with FPGA accelerator will be alternative to GPGPU solutions for real-time applications, especially in data centers or sever farms.

This paper proposes a statistical character structure modeling method. It represents each stroke by the distribution of the feature points. The character structure is represented by the joint distribution of the component strokes. In the proposed model, the stroke relationship is effectively reflected by the statistical dependency. It can represent all kinds of stroke relationship effectively in a systematic way. Based on the character representation, a stroke neighbor selection method is also proposed. It measures the importance of a stroke relationship by the mutual information among the strokes. With such a measure, the important neighbor relationships are selected by the nth order probability approximation method. The neighbor selection algorithm reduces the complexity significantly because we can reflect only some important relationships instead of all existing relationships. The proposed character modeling method was applied to a handwritten Chinese character recognition system. Applying a model-driven stroke extraction algorithm that cooperates with a selective matching algorithm, the proposed system is better than conventional structural recognition systems in analyzing degraded images. The effectiveness of the proposed methods was visualized by the experiments. The proposed method successfully detected and reflected the stroke relationships that seemed intuitively important. The overall recognition rate was 98.45 percent, which confirms the effectiveness of the proposed methods.

Structural matching and feature matching are two major streams in handwritten Chinese character recognition. The former is better in absorbing writing variation since it utilizes rich structural information from strokes directly. However, it is worse than the latter in handling noisy images, because of the difficulty in stroke extraction. For the ultimate performance in a recognition system, the structural information should be utilized without suffering from the faults of stroke extraction. This paper proposes a stroke-guided pixel matching method to satisfy the requirements. It uses a pixel-based character model to match images not depending on strokes. In order to utilize structural information, a stroke-guided pixel-grouping algorithm is also proposed. The initial experimental result shows that the proposed approach is promising.

Proposes some strategies to improve the recognition performance of a feature matching method for handwritten Chinese character recognition (HCCR). Favorable modifications are given to all stages throughout the recognition. In pre-processing, we devised a modified nonlinear normalization algorithm and a connectivity-preserving smoothing algorithm. For feature extraction, an efficient directional decomposition algorithm and a systematic approach to design a blurring mask are presented. Finally, a modified LVQ3 algorithm is applied to optimize the reference vectors for classification. The integrated effect of these strategies significantly improves the recognition performance. Recognition results on the large-vocabulary databases ETL8B2 and ETL9B are promising.