Yusuke Mukuta

Shusuke Takahama, Yusuke Kurose, Yusuke Mukuta et al.

Pathological image analysis is an important process for detecting abnormalities such as cancer from cell images. However, since the image size is generally very large, the cost of providing detailed annotations is high, which makes it difficult to apply machine learning techniques. One way to improve the performance of identifying abnormalities while keeping the annotation cost low is to use only labels for each slide, or to use information from another dataset that has already been labeled. However, such weak supervisory information often does not provide sufficient performance. In this paper, we proposed a new task setting to improve the classification performance of the target dataset without increasing annotation costs. And to solve this problem, we propose a pipeline that uses multiple instance learning (MIL) and domain adaptation (DA) methods. Furthermore, in order to combine the supervisory information of both methods effectively, we propose a method to create pseudo-labels with high confidence. We conducted experiments on the pathological image dataset we created for this study and showed that the proposed method significantly improves the classification performance compared to existing methods.

Bumjun Jung, Yusuke Mukuta, Tatsuya Harada

Time-series data analysis is important because numerous real-world tasks such as forecasting weather, electricity consumption, and stock market involve predicting data that vary over time. Time-series data are generally recorded over a long period of observation with long sequences owing to their periodic characteristics and long-range dependencies over time. Thus, capturing long-range dependency is an important factor in time-series data forecasting. To solve these problems, we proposed two novel modules, Grouped Self-Attention (GSA) and Compressed Cross-Attention (CCA). With both modules, we achieved a computational space and time complexity of order $O(l)$ with a sequence length $l$ under small hyperparameter limitations, and can capture locality while considering global information. The results of experiments conducted on time-series datasets show that our proposed model efficiently exhibited reduced computational complexity and performance comparable to or better than existing methods.

Yusuke Mori, Hiroaki Yamane, Yusuke Mukuta et al.

Storytelling has always been vital for human nature. From ancient times, humans have used stories for several objectives including entertainment, advertisement, and education. Various analyses have been conducted by researchers and creators to determine the way of producing good stories. The deep relationship between stories and emotions is a prime example. With the advancement in deep learning technology, computers are expected to understand and generate stories. This survey paper is intended to summarize and further contribute to the development of research being conducted on the relationship between stories and emotions. We believe creativity research is not to replace humans with computers, but to find a way of collaboration between humans and computers to enhance the creativity. With the intention of creating a new intersection between computational storytelling research and human creative writing, we introduced creative techniques used by professional storytellers.

Ryoma Kobayashi, Yusuke Mukuta, Tatsuya Harada

Learning from Label Proportions (LLP) is a weakly supervised learning method that aims to perform instance classification from training data consisting of pairs of bags containing multiple instances and the class label proportions within the bags. Previous studies on multiclass LLP can be divided into two categories according to the learning task: per-instance label classification and per-bag label proportion estimation. However, these methods often results in high variance estimates of the risk when applied to complex models, or lack statistical learning theory arguments. To address this issue, we propose new learning methods based on statistical learning theory for both per-instance and per-bag policies. We demonstrate that the proposed methods are respectively risk-consistent and classifier-consistent in an instance-wise manner, and analyze the estimation error bounds. Additionally, we present a heuristic approximation method that utilizes an existing method for regressing label proportions to reduce the computational complexity of the proposed methods. Through benchmark experiments, we demonstrated the effectiveness of the proposed methods.

Toshiki Katsube, Taiga Fukuhara, Kenichiro Ando et al.

This work addresses the scarcity of high-quality, large-scale resources for Japanese Vision-and-Language (V&L) modeling. We present a scalable and reproducible pipeline that integrates large-scale web collection with rigorous filtering/deduplication, object-detection-driven evidence extraction, and Large Language Model (LLM)-based refinement under grounding constraints. Using this pipeline, we build two resources: an image-caption dataset (DEJIMA-Cap) and a VQA dataset (DEJIMA-VQA), each containing 3.88M image-text pairs, far exceeding the size of existing Japanese V&L datasets. Human evaluations demonstrate that DEJIMA achieves substantially higher Japaneseness and linguistic naturalness than datasets constructed via translation or manual annotation, while maintaining factual correctness at a level comparable to human-annotated corpora. Quantitative analyses of image feature distributions further confirm that DEJIMA broadly covers diverse visual domains characteristic of Japan, complementing its linguistic and cultural representativeness. Models trained on DEJIMA exhibit consistent improvements across multiple Japanese multimodal benchmarks, confirming that culturally grounded, large-scale resources play a key role in enhancing model performance. All data sources and modules in our pipeline are licensed for commercial use, and we publicly release the resulting dataset and metadata to encourage further research and industrial applications in Japanese V&L modeling.

Yusuke Mukuta, Tatsuya Harada

This paper proposes a method to construct pretext tasks for self-supervised learning on group equivariant neural networks. Group equivariant neural networks are the models whose structure is restricted to commute with the transformations on the input. Therefore, it is important to construct pretext tasks for self-supervised learning that do not contradict this equivariance. To ensure that training is consistent with the equivariance, we propose two concepts for self-supervised tasks: equivariant pretext labels and invariant contrastive loss. Equivariant pretext labels use a set of labels on which we can define the transformations that correspond to the input change. Invariant contrastive loss uses a modified contrastive loss that absorbs the effect of transformations on each input. Experiments on standard image recognition benchmarks demonstrate that the equivariant neural networks exploit the proposed equivariant self-supervised tasks.

Motoki Omura, Yusuke Mukuta, Kazuki Ota et al.

Offline reinforcement learning (RL) aims to learn an optimal policy from a static dataset, making it particularly valuable in scenarios where data collection is costly, such as robotics. A major challenge in offline RL is distributional shift, where the learned policy deviates from the dataset distribution, potentially leading to unreliable out-of-distribution actions. To mitigate this issue, regularization techniques have been employed. While many existing methods utilize density ratio-based measures, such as the $f$-divergence, for regularization, we propose an approach that utilizes the Wasserstein distance, which is robust to out-of-distribution data and captures the similarity between actions. Our method employs input-convex neural networks (ICNNs) to model optimal transport maps, enabling the computation of the Wasserstein distance in a discriminator-free manner, thereby avoiding adversarial training and ensuring stable learning. Our approach demonstrates comparable or superior performance to widely used existing methods on the D4RL benchmark dataset. The code is available at https://github.com/motokiomura/Q-DOT .

Hiromichi Kamata, Yusuke Mukuta, Tatsuya Harada

Spiking neural networks (SNNs) can be run on neuromorphic devices with ultra-high speed and ultra-low energy consumption because of their binary and event-driven nature. Therefore, SNNs are expected to have various applications, including as generative models being running on edge devices to create high-quality images. In this study, we build a variational autoencoder (VAE) with SNN to enable image generation. VAE is known for its stability among generative models; recently, its quality advanced. In vanilla VAE, the latent space is represented as a normal distribution, and floating-point calculations are required in sampling. However, this is not possible in SNNs because all features must be binary time series data. Therefore, we constructed the latent space with an autoregressive SNN model, and randomly selected samples from its output to sample the latent variables. This allows the latent variables to follow the Bernoulli process and allows variational learning. Thus, we build the Fully Spiking Variational Autoencoder where all modules are constructed with SNN. To the best of our knowledge, we are the first to build a VAE only with SNN layers. We experimented with several datasets, and confirmed that it can generate images with the same or better quality compared to conventional ANNs. The code is available at https://github.com/kamata1729/FullySpikingVAE

Nabarun Goswami, Yusuke Mukuta, Tatsuya Harada

The success of models operating on tokenized data has heightened the need for effective tokenization methods, particularly in vision and auditory tasks where inputs are naturally continuous. A common solution is to employ Vector Quantization (VQ) within VQ Variational Autoencoders (VQVAEs), transforming inputs into discrete tokens by clustering embeddings in Euclidean space. However, Euclidean embeddings not only suffer from inefficient packing and limited separation - due to their polynomial volume growth - but are also prone to codebook collapse, where only a small subset of codebook vectors are effectively utilized. To address these limitations, we introduce HyperVQ, a novel approach that formulates VQ as a hyperbolic Multinomial Logistic Regression (MLR) problem, leveraging the exponential volume growth in hyperbolic space to mitigate collapse and improve cluster separability. Additionally, HyperVQ represents codebook vectors as geometric representatives of hyperbolic decision hyperplanes, encouraging disentangled and robust latent representations. Our experiments demonstrate that HyperVQ matches traditional VQ in generative and reconstruction tasks, while surpassing it in discriminative performance and yielding a more efficient and disentangled codebook.

Ryo Watanabe, Yusuke Mukuta, Tatsuya Harada

Spiking neural networks (SNNs) have garnered considerable attention owing to their ability to run on neuromorphic devices with super-high speeds and remarkable energy efficiencies. SNNs can be used in conventional neural network-based time- and energy-consuming applications. However, research on generative models within SNNs remains limited, despite their advantages. In particular, diffusion models are a powerful class of generative models, whose image generation quality surpass that of the other generative models, such as GANs. However, diffusion models are characterized by high computational costs and long inference times owing to their iterative denoising feature. Therefore, we propose a novel approach fully spiking denoising diffusion implicit model (FSDDIM) to construct a diffusion model within SNNs and leverage the high speed and low energy consumption features of SNNs via synaptic current learning (SCL). SCL fills the gap in that diffusion models use a neural network to estimate real-valued parameters of a predefined probabilistic distribution, whereas SNNs output binary spike trains. The SCL enables us to complete the entire generative process of diffusion models exclusively using SNNs. We demonstrate that the proposed method outperforms the state-of-the-art fully spiking generative model.

Motoki Omura, Takayuki Osa, Yusuke Mukuta et al.

In deep reinforcement learning, estimating the value function to evaluate the quality of states and actions is essential. The value function is often trained using the least squares method, which implicitly assumes a Gaussian error distribution. However, a recent study suggested that the error distribution for training the value function is often skewed because of the properties of the Bellman operator, and violates the implicit assumption of normal error distribution in the least squares method. To address this, we proposed a method called Symmetric Q-learning, in which the synthetic noise generated from a zero-mean distribution is added to the target values to generate a Gaussian error distribution. We evaluated the proposed method on continuous control benchmark tasks in MuJoCo. It improved the sample efficiency of a state-of-the-art reinforcement learning method by reducing the skewness of the error distribution.

Haruki Abe, Takayuki Osa, Yusuke Mukuta et al.

Scalable robot policy pre-training has been hindered by the high cost of collecting high-quality demonstrations for each platform. In this study, we address this issue by uniting offline reinforcement learning (offline RL) with cross-embodiment learning. Offline RL leverages both expert and abundant suboptimal data, and cross-embodiment learning aggregates heterogeneous robot trajectories across diverse morphologies to acquire universal control priors. We perform a systematic analysis of this offline RL and cross-embodiment paradigm, providing a principled understanding of its strengths and limitations. To evaluate this offline RL and cross-embodiment paradigm, we construct a suite of locomotion datasets spanning 16 distinct robot platforms. Our experiments confirm that this combined approach excels at pre-training with datasets rich in suboptimal trajectories, outperforming pure behavior cloning. However, as the proportion of suboptimal data and the number of robot types increase, we observe that conflicting gradients across morphologies begin to impede learning. To mitigate this, we introduce an embodiment-based grouping strategy in which robots are clustered by morphological similarity and the model is updated with a group gradient. This simple, static grouping substantially reduces inter-robot conflicts and outperforms existing conflict-resolution methods.

Haruo Fujiwara, Yusuke Mukuta, Tatsuya Harada

Recent advances in text-driven 3D scene editing and stylization, which leverage the powerful capabilities of 2D generative models, have demonstrated promising outcomes. However, challenges remain in ensuring high-quality stylization and view consistency simultaneously. Moreover, applying style consistently to different regions or objects in the scene with semantic correspondence is a challenging task. To address these limitations, we introduce techniques that enhance the quality of 3D stylization while maintaining view consistency and providing optional region-controlled style transfer. Our method achieves stylization by re-training an initial 3D representation using stylized multi-view 2D images of the source views. Therefore, ensuring both style consistency and view consistency of stylized multi-view images is crucial. We achieve this by extending the style-aligned depth-conditioned view generation framework, replacing the fully shared attention mechanism with a single reference-based attention-sharing mechanism, which effectively aligns style across different viewpoints. Additionally, inspired by recent 3D inpainting methods, we utilize a grid of multiple depth maps as a single-image reference to further strengthen view consistency among stylized images. Finally, we propose Multi-Region Importance-Weighted Sliced Wasserstein Distance Loss, allowing styles to be applied to distinct image regions using segmentation masks from off-the-shelf models. We demonstrate that this optional feature enhances the faithfulness of style transfer and enables the mixing of different styles across distinct regions of the scene. Experimental evaluations, both qualitative and quantitative, demonstrate that our pipeline effectively improves the results of text-driven 3D stylization.

Haruo Fujiwara, Yusuke Mukuta, Tatsuya Harada

We propose a simple yet effective pipeline for stylizing a 3D scene, harnessing the power of 2D image diffusion models. Given a NeRF model reconstructed from a set of multi-view images, we perform 3D style transfer by refining the source NeRF model using stylized images generated by a style-aligned image-to-image diffusion model. Given a target style prompt, we first generate perceptually similar multi-view images by leveraging a depth-conditioned diffusion model with an attention-sharing mechanism. Next, based on the stylized multi-view images, we propose to guide the style transfer process with the sliced Wasserstein loss based on the feature maps extracted from a pre-trained CNN model. Our pipeline consists of decoupled steps, allowing users to test various prompt ideas and preview the stylized 3D result before proceeding to the NeRF fine-tuning stage. We demonstrate that our method can transfer diverse artistic styles to real-world 3D scenes with competitive quality. Result videos are also available on our project page: https://haruolabs.github.io/style-n2n/

Motoki Omura, Takayuki Osa, Yusuke Mukuta et al.

In offline reinforcement learning, in-sample learning methods have been widely used to prevent performance degradation caused by evaluating out-of-distribution actions from the dataset. Extreme Q-learning (XQL) employs a loss function based on the assumption that Bellman error follows a Gumbel distribution, enabling it to model the soft optimal value function in an in-sample manner. It has demonstrated strong performance in both offline and online reinforcement learning settings. However, issues remain, such as the instability caused by the exponential term in the loss function and the risk of the error distribution deviating from the Gumbel distribution. Therefore, we propose Maclaurin Expanded Extreme Q-learning to enhance stability. In this method, applying Maclaurin expansion to the loss function in XQL enhances stability against large errors. This approach involves adjusting the modeled value function between the value function under the behavior policy and the soft optimal value function, thus achieving a trade-off between stability and optimality depending on the order of expansion. It also enables adjustment of the error distribution assumption from a normal distribution to a Gumbel distribution. Our method significantly stabilizes learning in online RL tasks from DM Control, where XQL was previously unstable. Additionally, it improves performance in several offline RL tasks from D4RL.

Ryo Umagami, Yu Ono, Yusuke Mukuta et al.

It is imperative to discern the relationships between multiple time series for accurate forecasting. In particular, for stock prices, components are often divided into groups with the same characteristics, and a model that extracts relationships consistent with this group structure should be effective. Thus, we propose the concept of hierarchical permutation-equivariance, focusing on index swapping of components within and among groups, to design a model that considers this group structure. When the prediction model has hierarchical permutation-equivariance, the prediction is consistent with the group relationships of the components. Therefore, we propose a hierarchically permutation-equivariant model that considers both the relationship among components in the same group and the relationship among groups. The experiments conducted on real-world data demonstrate that the proposed method outperforms existing state-of-the-art methods.

Yusuke Mori, Hiroaki Yamane, Ryohei Shimizu et al.

When humans write, they may unintentionally omit some information. Complementing the omitted information using a computer is helpful in providing writing support. Recently, in the field of story understanding and generation, story completion (SC) was proposed to generate the missing parts of an incomplete story. Although its applicability is limited because it requires that the user have prior knowledge of the missing part of a story, missing position prediction (MPP) can be used to compensate for this problem. MPP aims to predict the position of the missing part, but the prerequisite knowledge that "one sentence is missing" is still required. In this study, we propose Variable Number MPP (VN-MPP), a new MPP task that removes this restriction; that is, the task to predict multiple missing sentences or to judge whether there are no missing sentences in the first place. We also propose two methods for this new MPP task. Furthermore, based on the novel task and methods, we developed a creative writing support system, COMPASS. The results of a user experiment involving professional creators who write texts in Japanese confirm the efficacy and utility of the developed system.

Kohei Uehara, Yusuke Mori, Yusuke Mukuta et al.

Image narrative generation is a task to create a story from an image with a subjective viewpoint. Given the importance of the subjective feelings of writers, readers, and characters in storytelling, an image narrative generation method should consider human emotion. In this study, we propose a novel method of image narrative generation called ViNTER (Visual Narrative Transformer with Emotion arc Representation), which takes "emotion arc" as input to capture a sequence of emotional changes. Since emotion arcs represent the trajectory of emotional change, it is expected that we can include detailed information about the emotional changes in the story to the model. We present experimental results of both automatic and manual evaluations on the Image Narrative dataset and demonstrate the effectiveness of the proposed approach.

Yuki Kawana, Yusuke Mukuta, Tatsuya Harada

Articulated objects exist widely in the real world. However, previous 3D generative methods for unsupervised part decomposition are unsuitable for such objects, because they assume a spatially fixed part location, resulting in inconsistent part parsing. In this paper, we propose PPD (unsupervised Pose-aware Part Decomposition) to address a novel setting that explicitly targets man-made articulated objects with mechanical joints, considering the part poses. We show that category-common prior learning for both part shapes and poses facilitates the unsupervised learning of (1) part decomposition with non-primitive-based implicit representation, and (2) part pose as joint parameters under single-frame shape supervision. We evaluate our method on synthetic and real datasets, and we show that it outperforms previous works in consistent part parsing of the articulated objects based on comparable part pose estimation performance to the supervised baseline.

Sho Maeoki, Yusuke Mukuta, Tatsuya Harada

In this paper we undertake the task of text-based video moment retrieval from a corpus of videos. To train the model, text-moment paired datasets were used to learn the correct correspondences. In typical training methods, ground-truth text-moment pairs are used as positive pairs, whereas other pairs are regarded as negative pairs. However, aside from the ground-truth pairs, some text-moment pairs should be regarded as positive. In this case, one text annotation can be positive for many video moments. Conversely, one video moment can be corresponded to many text annotations. Thus, there are many-to-many correspondences between the text annotations and video moments. Based on these correspondences, we can form potentially relevant pairs, which are not given as ground truth yet are not negative; effectively incorporating such relevant pairs into training can improve the retrieval performance. The text query should describe what is happening in a video moment. Hence, different video moments annotated with similar texts, which contain a similar action, are likely to hold the similar action, thus these pairs can be considered as potentially relevant pairs. In this paper, we propose a novel training method that takes advantage of potentially relevant pairs, which are detected based on linguistic analysis about text annotation. Experiments on two benchmark datasets revealed that our method improves the retrieval performance both quantitatively and qualitatively.

Naoya Fushishita, Antonio Tejero-de-Pablos, Yusuke Mukuta et al.

Generating videos predicting the future of a given sequence has been an area of active research in recent years. However, an essential problem remains unsolved: most of the methods require large computational cost and memory usage for training. In this paper, we propose a novel method for generating future prediction videos with less memory usage than the conventional methods. This is a critical stepping stone in the path towards generating videos with high image quality, similar to that of generated images in the latest works in the field of image generation. We achieve high-efficiency by training our method in two stages: (1) image reconstruction to encode video frames into latent variables, and (2) latent variable prediction to generate the future sequence. Our method achieves an efficient compression of video into low-dimensional latent variables by decomposing each frame according to its hierarchical structure. That is, we consider that video can be separated into background and foreground objects, and that each object holds time-varying and time-independent information independently. Our experiments show that the proposed method can efficiently generate future prediction videos, even for complex datasets that cannot be handled by previous methods.

Yuki Kawana, Yusuke Mukuta, Tatsuya Harada

Reconstructing 3D objects from 2D images is a fundamental task in computer vision. Accurate structured reconstruction by parsimonious and semantic primitive representation further broadens its application. When reconstructing a target shape with multiple primitives, it is preferable that one can instantly access the union of basic properties of the shape such as collective volume and surface, treating the primitives as if they are one single shape. This becomes possible by primitive representation with unified implicit and explicit representations. However, primitive representations in current approaches do not satisfy all of the above requirements at the same time. To solve this problem, we propose a novel primitive representation named neural star domain (NSD) that learns primitive shapes in the star domain. We show that NSD is a universal approximator of the star domain and is not only parsimonious and semantic but also an implicit and explicit shape representation. We demonstrate that our approach outperforms existing methods in image reconstruction tasks, semantic capabilities, and speed and quality of sampling high-resolution meshes.

Ryohei Shimizu, Yusuke Mukuta, Tatsuya Harada

Hyperbolic spaces, which have the capacity to embed tree structures without distortion owing to their exponential volume growth, have recently been applied to machine learning to better capture the hierarchical nature of data. In this study, we generalize the fundamental components of neural networks in a single hyperbolic geometry model, namely, the Poincaré ball model. This novel methodology constructs a multinomial logistic regression, fully-connected layers, convolutional layers, and attention mechanisms under a unified mathematical interpretation, without increasing the parameters. Experiments show the superior parameter efficiency of our methods compared to conventional hyperbolic components, and stability and outperformance over their Euclidean counterparts.

Shusuke Takahama, Yusuke Kurose, Yusuke Mukuta et al.

Histopathological image analysis is an essential process for the discovery of diseases such as cancer. However, it is challenging to train CNN on whole slide images (WSIs) of gigapixel resolution considering the available memory capacity. Most of the previous works divide high resolution WSIs into small image patches and separately input them into the model to classify it as a tumor or a normal tissue. However, patch-based classification uses only patch-scale local information but ignores the relationship between neighboring patches. If we consider the relationship of neighboring patches and global features, we can improve the classification performance. In this paper, we propose a new model structure combining the patch-based classification model and whole slide-scale segmentation model in order to improve the prediction performance of automatic pathological diagnosis. We extract patch features from the classification model and input them into the segmentation model to obtain a whole slide tumor probability heatmap. The classification model considers patch-scale local features, and the segmentation model can take global information into account. We also propose a new optimization method that retains gradient information and trains the model partially for end-to-end learning with limited GPU memory capacity. We apply our method to the tumor/normal prediction on WSIs and the classification performance is improved compared with the conventional patch-based method.

Kosuke Arase, Yusuke Mukuta, Tatsuya Harada

Instance segmentation on 3D point clouds is one of the most extensively researched areas toward the realization of autonomous cars and robots. Certain existing studies have split input point clouds into small regions such as 1m x 1m; one reason for this is that models in the studies cannot consume a large number of points because of the large space complexity. However, because such small regions occasionally include a very small number of instances belonging to the same class, an evaluation using existing metrics such as mAP is largely affected by the category recognition performance. To address these problems, we propose a new method with space complexity O(Np) such that large regions can be consumed, as well as novel metrics for tasks that are independent of the categories or size of the inputs. Our method learns a mapping from input point clouds to an embedding space, where the embeddings form clusters for each instance and distinguish instances using these clusters during testing. Our method achieves state-of-the-art performance using both existing and the proposed metrics. Moreover, we show that our new metric can evaluate the performance of a task without being affected by any other condition.

Wataru Kawai, Yusuke Mukuta, Tatsuya Harada

Graphs are ubiquitous real-world data structures, and generative models that approximate distributions over graphs and derive new samples from them have significant importance. Among the known challenges in graph generation tasks, scalability handling of large graphs and datasets is one of the most important for practical applications. Recently, an increasing number of graph generative models have been proposed and have demonstrated impressive results. However, scalability is still an unresolved problem due to the complex generation process or difficulty in training parallelization. In this paper, we first define scalability from three different perspectives: number of nodes, data, and node/edge labels. Then, we propose GRAM, a generative model for graphs that is scalable in all three contexts, especially in training. We aim to achieve scalability by employing a novel graph attention mechanism, formulating the likelihood of graphs in a simple and general manner. Also, we apply two techniques to reduce computational complexity. Furthermore, we construct a unified and non-domain-specific evaluation metric in node/edge-labeled graph generation tasks by combining a graph kernel and Maximum Mean Discrepancy. Our experiments on synthetic and real-world graphs demonstrated the scalability of our models and their superior performance compared with baseline methods.

Yusuke Mukuta, Tatsuya Harada

In this study, a novel feature coding method that exploits invariance for transformations represented by a finite group of orthogonal matrices is proposed. We prove that the group-invariant feature vector contains sufficient discriminative information when learning a linear classifier using convex loss minimization. Based on this result, a novel feature model that explicitly consider group action is proposed for principal component analysis and k-means clustering, which are commonly used in most feature coding methods, and global feature functions. Although the global feature functions are in general complex nonlinear functions, the group action on this space can be easily calculated by constructing these functions as tensor-product representations of basic representations, resulting in an explicit form of invariant feature functions. The effectiveness of our method is demonstrated on several image datasets.

Yusuke Mukuta, Tatsuaki Machida, Tatsuya Harada

Covariance pooling is a feature pooling method with good classification accuracy. Because covariance features consist of second-order statistics, the scale of the feature elements are varied. Therefore, normalizing covariance features using a matrix square root affects the performance improvement. When pooling methods are applied to local features extracted from CNN models, the accuracy increases when the pooling function is back-propagatable and the feature-extraction model is learned in an end-to-end manner. Recently, the iterative polynomial approximation method for the matrix square root of a covariance feature was proposed, and resulted in a faster and more stable training than the methods based on singular-value decomposition. In this paper, we propose an extension of compact bilinear pooling, which is a compact approximation of the standard covariance feature, to the polynomials of the covariance feature. Subsequently, we apply the proposed approximation to the polynomial corresponding to the matrix square root to obtain a compact approximation for the square root of the covariance feature. Our method approximates a higher-dimensional polynomial of a covariance by the weighted sum of the approximate features corresponding to a pair of local features based on the similarity of the local features. We apply our method for standard fine-grained image recognition datasets and demonstrate that the proposed method shows comparable accuracy with fewer dimensions than the original feature.

Naoya Fushishita, Antonio Tejero-de-Pablos, Yusuke Mukuta et al.

Predicting future human behavior from an input human video is a useful task for applications such as autonomous driving and robotics. While most previous works predict a single future, multiple futures with different behavior can potentially occur. Moreover, if the predicted future is too short (e.g., less than one second), it may not be fully usable by a human or other systems. In this paper, we propose a novel method for future human pose prediction capable of predicting multiple long-term futures. This makes the predictions more suitable for real applications. Also, from the input video and the predicted human behavior, we generate future videos. First, from an input human video, we generate sequences of future human poses (i.e., the image coordinates of their body-joints) via adversarial learning. Adversarial learning suffers from mode collapse, which makes it difficult to generate a variety of multiple poses. We solve this problem by utilizing two additional inputs to the generator to make the outputs diverse, namely, a latent code (to reflect various behaviors) and an attraction point (to reflect various trajectories). In addition, we generate long-term future human poses using a novel approach based on unidimensional convolutional neural networks. Last, we generate an output video based on the generated poses for visualization. We evaluate the generated future poses and videos using three criteria (i.e., realism, diversity and accuracy), and show that our proposed method outperforms other state-of-the-art works.

Keisuke Hagiwara, Yusuke Mukuta, Tatsuya Harada

So far, research to generate captions from images has been carried out from the viewpoint that a caption holds sufficient information for an image. If it is possible to generate an image that is close to the input image from a generated caption, i.e., if it is possible to generate a natural language caption containing sufficient information to reproduce the image, then the caption is considered to be faithful to the image. To make such regeneration possible, learning using the cycle-consistency loss is effective. In this study, we propose a method of generating captions by learning end-to-end mutual transformations between images and texts. To evaluate our method, we perform comparative experiments with and without the cycle consistency. The results are evaluated by an automatic evaluation and crowdsourcing, demonstrating that our proposed method is effective.

Yusuke Mukuta, Akisato Kimura, David B Adrian et al.

The current state-of-the-art in feature learning relies on the supervised learning of large-scale datasets consisting of target content items and their respective category labels. However, constructing such large-scale fully-labeled datasets generally requires painstaking manual effort. One possible solution to this problem is to employ community contributed text tags as weak labels, however, the concepts underlying a single text tag strongly depends on the users. We instead present a new paradigm for learning discriminative features by making full use of the human curation process on social networking services (SNSs). During the process of content curation, SNS users collect content items manually from various sources and group them by context, all for their own benefit. Due to the nature of this process, we can assume that (1) content items in the same group share the same semantic concept and (2) groups sharing the same images might have related semantic concepts. Through these insights, we can define human curated groups as weak labels from which our proposed framework can learn discriminative features as a representation in the space of semantic concepts the users intended when creating the groups. We show that this feature learning can be formulated as a problem of link prediction for a bipartite graph whose nodes corresponds to content items and human curated groups, and propose a novel method for feature learning based on sparse coding or network fine-tuning.

Kuniaki Saito, Yusuke Mukuta, Yoshitaka Ushiku et al.

Obtaining common representations from different modalities is important in that they are interchangeable with each other in a classification problem. For example, we can train a classifier on image features in the common representations and apply it to the testing of the text features in the representations. Existing multi-modal representation learning methods mainly aim to extract rich information from paired samples and train a classifier by the corresponding labels; however, collecting paired samples and their labels simultaneously involves high labor costs. Addressing paired modal samples without their labels and single modal data with their labels independently is much easier than addressing labeled multi-modal data. To obtain the common representations under such a situation, we propose to make the distributions over different modalities similar in the learned representations, namely modality-invariant representations. In particular, we propose a novel algorithm for modality-invariant representation learning, named Deep Modality Invariant Adversarial Network (DeMIAN), which utilizes the idea of Domain Adaptation (DA). Using the modality-invariant representations learned by DeMIAN, we achieved better classification accuracy than with the state-of-the-art methods, especially for some benchmark datasets of zero-shot learning.