Masato Kiyama

Subgraph matching is a fundamental problem with applications across various domains, involving the identification of a query graph within a larger target graph. Graph Neural Networks (GNNs) are frequently employed to address this problem. In GNN-based subgraph matching, new node features are generated through convolutional operations that aggregate information from neighboring nodes. This process enables the identification of similar node pairs and facilitates the detection of subgraphs in the target graph that correspond to the query graph.Typically, GNN-based subgraph matching aims to identify subgraphs that perfectly match the query graph, often by using GNNs that share weights between the target and query graphs. However, real-world graph data often contain noise. Consequently, GNNs with shared weights may produce dissimilar node features after convolution, leading to incorrect matching. To address this issue, we propose TDGNet, a model designed for approximate subgraph matching on noisy graphs. TDGNet employs distinct graph convolutional networks for the query and target graphs to ensure the similarity of convolved node features while enhancing robustness to noise. Furthermore, TDGNet introduces a mechanism that concatenates the original node features of the query graph with information aggregated from the target graph via one-way cross-propagation, combined with an edge connectivity loss function, thereby improving the model’s structural understanding of the matched subgraphs. Experimental evaluations on three public datasets with injected noise demonstrate that TDGNet achieves accuracy comparable to or higher than existing learning-based subgraph matching methods. In particular, on the COX2 dataset with 10% noise, TDGNet achieved 91.9% matching accuracy, outperforming the existing method’s 89.9%.

Background: Catheter ablation for atypical atrial flutter (AFL) is challenging owing to its complex circuit. Consequently, the success rate varies depending on the circuit complexity and the operator's level of experience. Objective: This study aimed to create software using artificial intelligence to identify the critical isthmus of atypical AFL on 3D mapping. Methods: Raw data from 19 episodes of atypical AFL were extracted from the 3D mapping system. Local data, including location, activation timing, and voltage amplitude, were used to design software to identify areas where wavefronts converge and conduction slowing (which are characteristics of the optimal treatment area), that is, the critical isthmus, in patients with atypical AFL. Subsequently, the newly developed software was validated by evaluating the concordance between the software‐estimated critical isthmus location and the actual AFL termination area. Results: This newly developed software identified the estimated critical isthmus in all cases in 24.9 (6.7–46.2) seconds. The concordance rate between Top 1 prediction area and actual AFL termination area was high of 79% (95% confidence interval: 56.7%–91.5%). Although 4 of the 19 episodes did not show concordance, it is assumed that ablation at the estimated critical isthmus may have resulted in arrhythmia termination in two cases. Conclusion: This newly developed software identified the optimal treatment area for the atypical AFL with high accuracy. Use of this software may enable faster and more precise treatments. Furthermore, it may allow less experienced operators to achieve results that are comparable to those of experienced operators.

With increasing incidents due to congestion at events, effective pedestrian guidance has become a critical safety concern. Recent research has explored the application of reinforcement learning to crowd simulation, where agents learn optimal actions through trial and error to maximize rewards based on environmental states. This study investigates the use of reinforcement learning and simulation techniques to mitigate pedestrian congestion through improved guidance systems. We employ the Multi-Agent Deep Deterministic Policy Gradient (MA-DDPG), a multi-agent reinforcement learning approach, and propose an enhanced method for learning the Q-function for actors within the MA-DDPG framework. Using the Mojiko Fireworks Festival dataset as a case study, we evaluated the effectiveness of our proposed method by comparing congestion levels with existing approaches. The results demonstrate that our method successfully reduces congestion, with agents exhibiting superior cooperation in managing crowd flow. This improvement in agent coordination suggests the potential for practical applications in real-world crowd management scenarios.

Subgraph matching, which involves identifying a query graph within a larger data graph, is a fundamental problem that has applications in various fields. Although graph neural networks (GNNs) are commonly used in learning-based subgraph matching, they face challenges due to their convolution process. Specifically, GNNs generate new node features by aggregating information from adjacent nodes, meaning that even if nodes in the query and data graphs start with identical features, the presence of extraneous nodes or edges in the data graph can distort feature representations and hinder accurate matching. Ensuring feature similarity between the query and data graphs despite these extraneous elements is a key challenge. To address this, we propose ENDNet, a model that improves subgraph matching by detecting and mitigating the influence of extraneous nodes. ENDNet uses a denormalized matching matrix to identify extra nodes and neutralizes their impact by setting their feature values to zero before aggregation and convolution. Additionally, ENDNet incorporates a shared-graph convolutional network, leveraging the sigmoid function to refine feature extraction. Experiments on four open datasets demonstrate that ENDNet outperforms existing trainable subgraph matching models, improving accuracy from 91.6% to 99.1% on the COX2 dataset. Ablation studies further confirm the effectiveness of extra-node determination and the shared-graph convolutional network, highlighting their importance in subgraph matching. The code for ENDNet is available on GitHub.

Background: Conventional medical image retrieval methods treat images and text as independent embeddings, limiting their ability to fully utilize the complementary information from both modalities. This separation often results in suboptimal retrieval performance, as the intricate relationships between images and text remain underexplored. Methods: To address this limitation, we propose a novel retrieval method that integrates medical image and text embeddings using a cross-attention mechanism. Our approach creates a unified representation by directly modeling the interactions between the two modalities, significantly enhancing retrieval accuracy. Results: Built upon the pre-trained BioMedCLIP model, our method outperforms existing techniques across multiple metrics, achieving the highest mean Average Precision (mAP) on the MIMIC-CXR dataset. Conclusions: These results highlight the effectiveness of our method in advancing multimodal medical image retrieval and set the stage for further innovation in the field.

Multi-camera people tracking (MCPT) has become increasingly relevant in recent years as the demand for accurate people-tracking systems has increased along with the growing number of surveillance cameras. There are challenges that MCPT must overcome, such as changes in a person’s appearance due to illumination and changes in viewpoint and posture between cameras. Additionally, occlusion is likely to occur in crowded places like tourist spots, making it even more difficult to perform accurate tracking based on a person’s appearance. To address these problems, we propose an MCPT system that uses additional spatio-temporal and group information. First, the system uses spatio-temporal filtering to remove candidates that are considered irrelevant. Then, group-aware matching is used to correct ID matching errors based solely on the features of an individual’s appearance. In this paper, we evaluate this system on data collected from surveillance cameras at Kumamoto Castle, a tourist spot designated as a National Important Cultural Property. This dataset contains images of people in a wide age range and with various degrees of crowding. We demonstrate that our system is effective in scenarios with large crowds, and that the additional contextual information is helpful when it is difficult to track based on a person’s appearance alone.

The design of a lightweight deep learning model would be an ideal solution for overcoming resource limitations when implementing artificial intelligence in edge sites. In this study, we propose a lightweight deep neural network that uses a Mixer-type architecture based on nonlinear vector autoregression (NVAR), which we refer to as Mixer-type NVAR. We applied overlapping patch embedding to enrich the image input and Sequencer architecture for vertical and horizontal operation inside the Mixer-type NVAR. We utilized a window partition technique and general quadratic positional encoding to increase the performance of the proposed model. Our model achieved a top-1 accuracy of 82.48 for the CIFAR-10 dataset with 0.159Mparameters and 98.36 for MNIST with 0.106Mparameters. Moreover, we evaluated its throughput on a central processing unit, which was 190.1 images per second for CIFAR-10 and 106.7 images per second for the MNIST dataset. These results are competitive with the state-of-the-art convolutional neural network-based model, MLP-Mixer, and the traditional reservoir-computing-based Mixer model with the same tuning of hyperparameters.

In this paper, we perform the super-resolution of sea surface temperature data with the enhanced super-resolution generative adversarial network (ESRGAN), which is a deep neural network-based single-image super-resolution (SISR) method that uses a generative adversarial network (GAN). We generate high-quality super-resolution data with ESRGAN and with the super-resolution convolutional neural network (SRCNN) and residual-in-residual dense block network (RRDBNet) methods, which are based on convolutional neural networks (CNNs). The images generated with these methods are compared with high-resolution optimum interpolation sea surface temperature (OISST) data using root mean square error (RMSE), learned perceptual image patch similarity (LPIPS), and perceptual index (PI) evaluation methods. RRDBNet has a better RMSE than SRCNN and ESRGAN. However, CNN-based SISR methods do not provide a faithful representation of the ocean currents of OISST. ESRGAN has a better LPIPS and PI than CNN-based methods and can represent the complex distribution of ocean currents.