| YANG YILING | D, 中間発表 | コンピューティング・アーキテクチャ | 中島 康彦, | 林 優一, | 張 任遠, | KAN Yirong, | PHAM HOAI LUAN, | Le Vu Trung Duong |
|
title: Deep Variational Information Bottleneck with Gaussian Latent Representations for Spiking Neural Networks
abstract: Spiking neural networks (SNNs) are promising for energy-efficient computing because they process information through discrete spike events. However, improving the generalization and robustness of directly trained SNNs remains challenging. This study develops a deep variational information bottleneck (DVIB) framework for SNNs that balances predictive information preservation and input information compression. The task-relevance objective is implemented using categorical cross-entropy, while the compression objective is upper-bounded by the Kullback-Leibler divergence between the stochastic encoder distribution and a standard Gaussian prior. The stochastic bottleneck is introduced at the synaptic input currents before spike generation and trained end-to-end using backpropagation through time, surrogate gradients, and the reparameterization trick. To model the stochastic latent representation, two Gaussian encoder distributions are considered: a diagonal covariance model that treats latent components as conditionally independent, and a Kronecker-structured covariance model that captures temporal dependencies together with feature-wise variance structure. The framework is evaluated on CIFAR-10 and CIFAR-100 classification using a spiking ResNet-19, and on DDD17 and DSEC event-based semantic segmentation using a SpikingResUNet. language of the presentation: English | ||||||||
| 安藤 拓翔 | M, 2回目発表 | コンピューティング・アーキテクチャ | 中島 康彦, | 林 優一, | 張 任遠, | KAN Yirong, | PHAM HOAI LUAN, | Le Vu Trung Duong |
|
title: AI Application Implementation and Software-Stack Improvements for Energy-Efficient Workloads on a CGLA
abstract: This reseach AI workload execution on IMAX, a CPU-Grounded Linear Array (CGLA). The goal is to reduce energy cost while preserving the execution flow of existing open-source AI applications. GPUs provide high throughput for LLM inference, speech recognition, and image generation, but power consumption is high. Fixed-function accelerators can reduce energy for narrow targets, but they are difficult to adapt to changing model structures and quantization formats. The work implements IMAX execution paths for llama.cpp, whisper.cpp, and stable-diffusion.cpp. The scope includes quantized and FP16 matrix operations, offload boundaries, data-transfer policy, host-side control, CPU fallback, and phase-wise profiling. The LLM evaluation finds that decode execution on CGLA is constrained by LOAD, DRAIN, and host-side overhead rather than arithmetic alone. Under a 28 nm ASIC projection, the evaluated IMAX-based design improves PDP by up to 44.4x and EDP by up to 11.5x over an RTX 4090 for representative workloads. Q-Snap improves prompt-processing throughput by 1.62x. DRAIN-path restructuring raises transfer rate from 0.264 Gbps to 1.077 Gbps for Qwen3-0.6B Q8_0. The results support a software-stack view of CGLA acceleration. language of the presentation: English | ||||||||
| 岡本 光 | M, 2回目発表 | コンピューティング・アーキテクチャ | 中島 康彦, | 林 優一, | 張 任遠, | KAN Yirong, | PHAM HOAI LUAN, | Le Vu Trung Duong |
|
title: A High-Throughput 384-bit Montgomery Multiplier on an FPGA for zk-SNARKs
abstract: Montgomery multiplication is a dominant operation in zk-SNARK multi-scalar multiplication, and its throughput on an edge FPGA depends on both arithmetic efficienfy and data delivery. This research work is focused on Montgomery multiplier System-on-Chip (SoC), with newly introduced 128-bit Direct Memory Access (DMA) bus, dual-buffered Block RAM (BRAM), and optimal number of Montgomery multiplication cores. Compared to a baseline with 64-bit bus, single-buffered BRAM, and a single core, I/O latency is expected to be reduced from 1936 to 1168 cycles. Achievable throughput of Montgomery multiplication is estimated to be 1.99 times, while the latency of the core is not changed. On a KV260 FPGA board, its theoretical thoughput is 6.75 Gbps and the SoC outputs a result case every 14.21 clock cycles on average. A practical implementation and verification of the improved design is in progress, which broaden the application of zk-SNARKs in resource constrained edge devices. language of the presentation: English | ||||||||
| 堤 悠太 | M, 2回目発表 | コンピューティング・アーキテクチャ | 中島 康彦, | 林 優一, | 張 任遠, | KAN Yirong, | PHAM HOAI LUAN, | Le Vu Trung Duong |
|
title: *** Implementation of CNN for medical signal processing on FPGA ***
abstract: *** Continuous ECG monitoring on edge devices requires high detection accuracy under strict power and hardware resource constraints. FPGA-based implementations provide high power and area efficiency while enabling low-latency processing. Therefore, this work focuses on implementing an efficient CNN-based ECG analysis system on the KV260 FPGA platform. Wearable electrocardiogram (ECG) monitoring systems require hardware architectures with high area and power efficiency. Previous work proposed a convolutional neural network (CNN) architecture optimized for wearable ECG monitoring. In this study, we implement and evaluate the proposed CNN architecture on the Xilinx KV260 FPGA platform. As a software baseline for FPGA implementation, the software model described in the previous work was reproduced and evaluated. ECG signals were preprocessed using the same procedure as the previous work, and QRS detection was performed using the proposed CNN. The reproduced software model was evaluated in terms of sensitivity (Se) and positive predictivity (P+), and the results confirmed that the model was successfully validated and could be used as the reference model for FPGA implementation. For the hardware implementation, the first convolutional layer (Conv1D) of the CNN was implemented as a hardware IP on the KV260 FPGA. A reusable ALU-based architecture proposed in the previous work was adopted to share arithmetic resources and improve area efficiency. FPGA resource utilization and timing characteristics were evaluated, and the implementation satisfied the timing constraints. The implementation of the remaining CNN layers is currently in progress. *** language of the presentation: *** English *** | ||||||||
| 清水 瑛音 | M, 2回目発表 | 光メディアインタフェース | 向川 康博, | 清川 清, | 藤村 友貴, | 北野 和哉 | |
|
title: Diffusion-Based Separation of Mixed Images from Two Adhered Photographic Plates
abstract: Due to the effects of flooding and other factors, photographic plates have become stuck together and fused. In this study, I leverage the powerful prior knowledge of a base model and use a diffusion model to separate the blended image of two fused photographic plates. While separating the blended image constitutes an ill-posed problem, I propose a separation method that uses the reflection image from a black background plate as a constraint to solve this issue. Furthermore, by introducing Diffusion Posterior Sampling (DPS), I suppress hallucinations—a common challenge with diffusion models—thereby preserving the plates' value as historical documents. This method has demonstrated consistent separation performance in quantitative evaluations (L1/SSIM), confirming its fundamental effectiveness. Moving forward, I aim to further improve separation accuracy for fine-scale structures. language of the presentation: Japanese 発表題目: 拡散モデルを用いた貼り付いた2枚の写真乾板の混合画像分離 発表概要: 水害等の影響により、写真乾板同士が貼り付き、固着する問題が発生している。本研究では基盤モデルの強力な事前知識を活用し、拡散モデルを用いて、貼り付いた2枚の写真乾板の混合画像を分離する。混合画像からの分離は不良設定問題となるが、これを解決するため、黒天板による反射画像を制約とした分離手法を提案する。さらに、Diffusion Posterior Sampling (DPS) を導入することで拡散モデルの課題であるハルシネーションを抑制し、歴史的史料としての価値を維持する。本手法により、定量評価（L1/SSIM）において一定の分離性能が実証され、手法の基礎的な有効性が確認された。今後は、微細な構造におけるさらなる分離精度の向上を目指す。 | |||||||
| 森近 拓生 | M, 2回目発表 | 光メディアインタフェース | 向川 康博, | 清川 清, | 藤村 友貴, | 北野 和哉 | |
|
title: Transferring Scene Features for Robust 3D Reconstruction across Lighting Environment
abstract: 3D reconstruction, which recovers 3D shapes from images, is a task that has undergone significant advancements in the field of computer vision in recent years. However, existing models face limitations when performing reconstruction using images with different illumination environments, and challenges still remain. Therefore, this study aims to establish a robust and versatile method capable of high-quality 3D reconstruction even from a small number of images captured under varying illumination conditions. Specifically, we propose a method that unifies the illumination conditions of images via Image-to-Image translation using a diffusion model, while maintaining structural consistency. As our progress to date, we have devised a diffusion model-based method that uses depth images as a conditioning input—serving as an intermediate representation containing only pure shape information—while simultaneously learning illumination embeddings to represent the lighting environment, along with its evaluation methodology. language of the presentation: Japanese 発表題目: 照明環境の異なる画像に対する三次元再構成モデルの性能改善 発表概要: 画像から三次元形状を復元する三次元再構成は，コンピュータビジョン分野において近年著しい発展を遂げているタスクである。しかし，既存モデルは，照明環境が異なる画像を用いた再構成に制限があり，依然として課題が残されている。そこで本研究では，異なる照明環境下で撮影された少数の画像からでも，高品質な三次元再構成が可能な，ロバストで汎用性の高い手法の確立を目指す。具体的には，照明環境の異なる画像に対し，拡散モデルを用いたImage-to-Image変換により，構造的な一貫性を保持しつつ画像の照明環境を統一する手法を提案する。現在までの進捗として，純粋な形状情報のみを持つ中間表現として深度画像を条件付けに用い，同時に照明環境を表現する照明埋め込みを学習する拡散モデルの手法，およびその評価手法について考案した。 | |||||||
| MIN ZIYU | D, 中間発表 | サイバネティクス・リアリティ工学 | 清川 清, | 向川 康博, | 内山 英昭, | Perusquia Hernandez Monica, | 平尾 悠太朗 |
|
title: Stable Foveated Rendering for 3D Gaussian Splatting
abstract: 3D Gaussian Splatting (3DGS) enables real-time novel-view rendering, but virtual reality demands high-resolution stereo images at high refresh rates. This work develops a training-free foveation layer for pretrained 3DGS models. It combines gaze-contingent log-polar selection, persistent hash ranks, and camera-independent anchors to reduce rendering work, suppress gaze-induced selection popping, and preserve peripheral support. Across 13 scenes from three public datasets, the layer provides consistent GPU rendering speedups on held-out test views. At a matched selection budget, it reduces gaze-motion temporal error by more than 90% relative to per-frame reselection, while image quality near the gaze point remains close to full rendering. Equal-budget controls show that anchor placement reduces peripheral coverage artifacts. Peripheral degradation remains visible at the evaluated aggressive operating point, so the results define a configurable quality-performance-stability frontier. Future work will evaluate natural gaze traces and human perception in a head-mounted display. language of the presentation: English | |||||||
| YU XINMENG | M, 2回目発表 | 光メディアインタフェース（マルチモーダル環境認識） | 向川 康博☆, | 川西 康友, | 薗頭 元春, | 藤村 友貴, | 北野 和哉 |
|
title: High-Speed SLAM Using Mixture of Experts and Instance Information
abstract: In Visual SLAM, image matching is one of the most important fundamental components, and its results greatly influence the accuracy of pose estimation and map construction. However, existing matching methods still limit the performance of Visual SLAM. Sparse matching methods with higher speed tend to produce fewer matching points than dense methods, making it difficult to achieve both high speed and high accuracy. To address this problem, this study proposes dynamically switching matching methods by using a mixture of experts framework. Sparse matching is used in high-texture regions, where it can obtain a sufficient number of matching points, while semi-dense matching is used in low-texture regions. This reduces the frequency of semi-dense matching and improves speed while maintaining accuracy. Semi-dense matching is more accurate than sparse matching and still faster than dense matching commonly used in SLAM, but it has the drawback that its accuracy decreases due to mismatches. To address this issue, we use instance segmentation to assign IDs to objects and introduce a bonus–penalty rule based on ID consistency into the matching process of a semi-dense matching model. This allows the model to distinguish objects that look the same but differ in ID, thereby reducing mismatching. language of the presentation: Japanese 発表題目: 混合エキスパートとインスタンス情報を用いた高速な SLAM 発表概要: Visual SLAM では，画像マッチングが最も重要な基礎の一つであり，その結果は姿勢推定や地図構築の精度に大きく影響する．しかし，現在のマッチング方法は依然として Visual SLAM の性能を制限している．スパースマッチングはデンスマッチングより高速である一方，得られる対応点数が少なくなる傾向があるため，高速性と高精度を同時に達成することが難しい．この問題を解決するために，本研究では混合エキスパートの枠組みを用いて，状況に応じて最適なマッチング方法を切り替えることを提案する．十分な対応点数を得られる高テクスチャ領域ではスパースマッチングを用い，低テクスチャ領域ではセミデンスマッチングを用いる．これによって，セミデンスマッチングの使用頻度を下げ，精度を保ちながら速度の向上を目指す．セミデンスマッチングは，スパースマッチングより高精度であり，かつ SLAM で一般的に用いられるデンスマッチングよりも高速であるが，誤マッチにより精度が低下する難点がある．この困難に対処するため，インスタンスセグメンテーションを用いて物体に ID を付与し，セミデンスマッチングモデルのマッチング過程にIDの一貫性に基づくボーナスとペナルティのルールを加える．これによって，外観が同じでもIDの異なる物体をモデルが区別できるようにし，誤マッチの発生を抑える． | |||||||
| 新江田 航大 | D, 中間発表 | インタラクティブメディア設計学 | 加藤 博一, | 和田 隆広, | 澤邊 太志, | Isidro Butaslac, | 神原 誠之 |
|
title: A Pain Relief System Integrating Affective Touch and Agent Audio-Visual Stimulation
abstract:With the increasing prevalence of chronic pain, there is a growing demand to establish non-pharmacological analgesic approaches that do not rely on medication. This study aims to develop a multimodal pain relief system that alleviates pain by maximizing pleasant emotions through the integration of robotic affective touch (stroking at 1-10 cm/s) and audio-visual stimulation from a virtual agent. Specifically, the research will be conducted in three phases. First, we will present a virtual agent during the stroking motion and quantify the conditions under which audio-visual factors amplify pleasant emotions using a multiple regression model. Second, we will elucidate the degree to which the induced pleasant emotions suppress the temporal summation of pain at the spinal level, thereby deriving the physiological thresholds of the descending pain inhibitory system. Third, based on the obtained models, we will develop and validate a system that automatically generates optimal stimulation parameters adapted to individual sensitivities. The originality of this research lies in the integration of a dual-pathway pain inhibition mechanism, synchronizing spinal-level blockade via C-tactile (CT) fiber stimulation with top-down inhibition from higher brain centers via VR. This system is expected to be implemented in society as a non-pharmacological pain management technology in medical and nursing care settings. language of the presentation: Japanese | |||||||