Tatsuhito Hasegawa

Wenmin Li, Shunsuke Sakai, Zhongkai Zhao et al.

Diffusion-based dataset distillation has recently emerged as a promising paradigm for condensing large-scale datasets into compact synthetic sets. By leveraging pretrained generative priors, these methods can produce realistic class-conditional samples more efficiently than traditional matching-based approaches. However, most existing diffusion-based methods still adopt a rigid ``Generate-and-Use'' strategy, where the generated samples are directly treated as the final distilled set under a fixed images-per-class budget. Such a design tightly couples candidate generation with final budget allocation, which may result in redundant waste of the limited budget or insufficiently informative samples. In this paper, we propose ``Pool-Select-Refine'', a two-stage framework for allocation-aware generative dataset distillation. First, instead of directly using a fixed number of generated samples, we construct an over-complete candidate pool and select a compact subset under the target budget. Second, we refine the selected samples in latent space using soft-label supervision derived from the teacher model, improving semantic alignment while preserving the generative prior. This design explicitly decouples generation, selection, and refinement, enabling more effective use of the distillation budget. Experiments on large-scale and fine-grained image classification benchmarks show that the proposed framework delivers consistent gains over diffusion-based baselines. The results suggest that introducing a curation stage before refinement is a simple yet effective way to improve diffusion-based dataset distillation.

Tatsuhito Hasegawa, Kazuma Kondo

Sensor-based human activity recognition (HAR) is a paramount technology in the Internet of Things services. HAR using representation learning, which automatically learns a feature representation from raw data, is the mainstream method because it is difficult to interpret relevant information from raw sensor data to design meaningful features. Ensemble learning is a robust approach to improve generalization performance; however, deep ensemble learning requires various procedures, such as data partitioning and training multiple models, which are time-consuming and computationally expensive. In this study, we propose Easy Ensemble (EE) for HAR, which enables the easy implementation of deep ensemble learning in a single model. In addition, we propose input masking as a method for diversifying the input for EE. Experiments on a benchmark dataset for HAR demonstrated the effectiveness of EE and input masking and their characteristics compared with conventional ensemble learning methods.

Tatsuhito Hasegawa, Taisei Tanaka

Single-model ensembles (SMEs) have attracted attention as a way to approximate some of the benefits of deep ensembles within a single network. However, under an approximately matched parameter budget, it remains unclear whether model capacity should be concentrated in a single wide pathway or redistributed into many narrow and independent members. We investigate this question through the Multi-Narrow (MN) transformation, which converts a baseline CNN into an SME of narrow, path-wise independent branches while approximately preserving the dominant parameter budget. We systematically compare Single-Wide and Multi-Narrow configurations across different training-data regimes, architectures, and datasets. The results show that the effectiveness of MN is strongly data-dependent: weakly partitioned or baseline-wide models are preferable in data-rich settings, whereas highly partitioned MN models consistently outperform the baseline in low-data settings. This tendency is reproduced across multiple CNN architectures and image-classification datasets, suggesting that it is not specific to a single benchmark or model family. Analysis of internal representations shows that high-MN models learn more diverse and less redundant path-wise features. In low-data regimes, this diversity is broadly utilized and improves generalization, whereas in data-rich regimes, training becomes imbalanced and prediction is dominated by a small subset of paths. These findings clarify when and why Multi-Narrow transformation is effective, and provide practical guidance for allocating model capacity between width and member multiplicity under a limited budget.

Xiangteng He, Shunsuke Sakai, Kun Yuan et al.

Image-based Joint-Embedding Predictive Architecture (I-JEPA) learns visual representations by predicting latent embeddings of masked regions from visible context. However, it treats all regions uniformly and independently, lacking an explicit notion of where or in what order predictions should be made. Inspired by human visual perception, which deploys attention selectively and sequentially from the most informative to secondary regions, we propose DSeq-JEPA, a Discriminative Sequential Joint-Embedding Predictive Architecture that bridges predictive and autoregressive self-supervised learning, integrating JEPA-style latent prediction with GPT-style sequential reasoning. Specifically, DSeq-JEPA (i) first identifies primary discriminative regions based on a transformer-derived saliency map, emphasizing the distribution of visual importance, and then (ii) predicts subsequent regions in this discriminative order, progressively forming a curriculum-like semantic progression from primary to secondary cues -- a form of GPT-style pre-training. Extensive experiments across diverse tasks, including image classification (ImageNet), fine-grained visual categorization (iNaturalist21, CUB-200-2011, Stanford-Cars), detection and segmentation (MS-COCO, ADE20K), and low-level reasoning tasks (Clevr/Count, Clevr/Dist), demonstrate that DSeq-JEPA consistently focuses on more discriminative and generalizable representations than I-JEPA variants. Project page: https://github.com/SkyShunsuke/DSeq-JEPA.

Shunsuke Sakai, Shunsuke Tsuge, Tatsuhito Hasegawa

Neural network ensembles is a simple yet effective approach for enhancing generalization capabilities. The most common method involves independently training multiple neural networks initialized with different weights and then averaging their predictions during inference. However, this approach increases training time linearly with the number of ensemble members. To address this issue, we propose the novel ``\textbf{Noisy Deep Ensemble}'' method, significantly reducing the training time required for neural network ensembles. In this method, a \textit{parent model} is trained until convergence, and then the weights of the \textit{parent model} are perturbed in various ways to construct multiple \textit{child models}. This perturbation of the \textit{parent model} weights facilitates the exploration of different local minima while significantly reducing the training time for each ensemble member. We evaluated our method using diverse CNN architectures on CIFAR-10 and CIFAR-100 datasets, surpassing conventional efficient ensemble methods and achieving test accuracy comparable to standard ensembles. Code is available at \href{https://github.com/TSTB-dev/NoisyDeepEnsemble}{https://github.com/TSTB-dev/NoisyDeepEnsemble}

Tatsuhito Hasegawa

Human activity recognition (HAR) in Internet of Things (IoT) environments must cope with heterogeneous sensor settings that vary across datasets, devices, body locations, sensing modalities, and channel compositions. This heterogeneity makes conventional channel-fixed models difficult to reuse across sensing environments because their input representations are tightly coupled to predefined channel structures. To address this problem, we investigate strict channel-free HAR, in which a single shared model performs inference without assuming a fixed number, order, or semantic arrangement of input channels, and without relying on sensor-specific input layers or dataset-specific channel templates. We argue that fusion design is the central issue in this setting. Accordingly, we propose a channel-free HAR framework that combines channel-wise encoding with a shared encoder, metadata-conditioned late fusion via conditional batch normalization, and joint optimization of channel-level and fused predictions through a combination loss. The proposed model processes each channel independently to handle varying channel configurations, while sensor metadata such as body location, modality, and axis help recover structural information that channel-independent processing alone cannot retain. In addition, the joint loss encourages both the discriminability of individual channels and the consistency of the final fused prediction. Experiments on PAMAP2, together with robustness analysis on six HAR datasets, ablation studies, sensitivity analysis, efficiency evaluation, and cross-dataset transfer learning, demonstrate three main findings...

Wenmin Li, Shunsuke Sakai, Tatsuhito Hasegawa

Deploying machine learning models in resource-constrained environments, such as edge devices or rapid prototyping scenarios, increasingly demands distillation of large datasets into significantly smaller yet informative synthetic datasets. Current dataset distillation techniques, particularly Trajectory Matching methods, optimize synthetic data so that the model's training trajectory on synthetic samples mirrors that on real data. While demonstrating efficacy on medium-scale synthetic datasets, these methods fail to adequately preserve semantic richness under extreme sample scarcity. To address this limitation, we propose a novel dataset distillation method integrating contrastive learning during image synthesis. By explicitly maximizing instance-level feature discrimination, our approach produces more informative and diverse synthetic samples, even when dataset sizes are significantly constrained. Experimental results demonstrate that incorporating contrastive learning substantially enhances the performance of models trained on very small-scale synthetic datasets. This integration not only guides more effective feature representation but also significantly improves the visual fidelity of the synthesized images. Experimental results demonstrate that our method achieves notable performance improvements over existing distillation techniques, especially in scenarios with extremely limited synthetic data.

Tatsuhito Hasegawa, Shunsuke Sakai

In deep learning-based classification tasks, the softmax function's temperature parameter $T$ critically influences the output distribution and overall performance. This study presents a novel theoretical insight that the optimal temperature $T^*$ is uniquely determined by the dimensionality of the feature representations, thereby enabling training-free determination of $T^*$. Despite this theoretical grounding, empirical evidence reveals that $T^*$ fluctuates under practical conditions owing to variations in models, datasets, and other confounding factors. To address these influences, we propose and optimize a set of temperature determination coefficients that specify how $T^*$ should be adjusted based on the theoretical relationship to feature dimensionality. Additionally, we insert a batch normalization layer immediately before the output layer, effectively stabilizing the feature space. Building on these coefficients and a suite of large-scale experiments, we develop an empirical formula to estimate $T^*$ without additional training while also introducing a corrective scheme to refine $T^*$ based on the number of classes and task complexity. Our findings confirm that the derived temperature not only aligns with the proposed theoretical perspective but also generalizes effectively across diverse tasks, consistently enhancing classification performance and offering a practical, training-free solution for determining $T^*$.

Shunsuke Sakai, Xiangteng He, Chunzhi Gu et al.

Despite the remarkable success, recent reconstruction-based anomaly detection (AD) methods via diffusion modeling still involve fine-grained noise-strength tuning and computationally expensive multi-step denoising, leading to a fundamental tension between fidelity and efficiency. In this paper, we propose a novel inversion-based AD approach - detection via noising in latent space - which circumvents explicit reconstruction. Importantly, we contend that the limitations in prior reconstruction-based methods originate from the prevailing detection via denoising in RGB space paradigm. To address this, we model AD under a reconstruction-free formulation, which directly infers the final latent variable corresponding to the input image via DDIM inversion, and then measures the deviation based on the known prior distribution for anomaly scoring. Specifically, in approximating the original probability flow ODE using the Euler method, we only enforce very few inversion steps to noise the clean image to pursue inference efficiency. As the added noise is adaptively derived with the learned diffusion model, the original features for the clean testing image can still be leveraged to yield high detection accuracy. We perform extensive experiments and detailed analysis across three widely used image AD datasets under the unsupervised unified setting to demonstrate the effectiveness of our model, regarding state-of-the-art AD performance, and about 2 times inference time speedup without diffusion distillation.

Tatsuhito Hasegawa

In the research field of activity recognition, although it is difficult to collect a large amount of measured sensor data, there has not been much discussion about data augmentation (DA). In this study, I propose Octave Mix as a new synthetic-style DA method for sensor-based activity recognition. Octave Mix is a simple DA method that combines two types of waveforms by intersecting low and high frequency waveforms using frequency decomposition. In addition, I propose a DA ensemble model and its training algorithm to acquire robustness to the original sensor data while remaining a wide variety of feature representation. I conducted experiments to evaluate the effectiveness of my proposed method using four different benchmark datasets of sensing-based activity recognition. As a result, my proposed method achieved the best estimation accuracy. Furthermore, I found that ensembling two DA strategies: Octave Mix with rotation and mixup with rotation, make it possible to achieve higher accuracy.

Tatsuhito Hasegawa

There is a research field of human activity recognition that automatically recognizes a user's physical activity through sensing technology incorporated in smartphones and other devices. When sensing daily activity, various measurement conditions, such as device type, possession method, wearing method, and measurement application, are often different depending on the user and the date of the measurement. Models that predict activity from sensor values are often implemented by machine learning and are trained using a large amount of activity-labeled sensor data measured from many users who provide labeled sensor data. However, collecting activity-labeled sensor data using each user's individual smartphones causes data being measured in inconsistent environments that may degrade the estimation accuracy of machine learning. In this study, I propose an activity recognition method that is robust to different sampling rates -- even in the measurement environment. The proposed method applies an adversarial network and data augmentation by downsampling to a common activity recognition model to achieve the acquisition of feature representations that make the sampling rate unspecifiable. Using the Human Activity Sensing Consortium (HASC), which is a dataset of basic activity recognition using smartphone sensors, I conducted an evaluation experiment to simulate an environment in which various sampling rates were measured. As a result, I found that estimation accuracy was reduced by the conventional method in the above environment and could be improved by my proposed method.