As the scales of neural networks increase, techniques that enable them to run with low computational cost and energy efficiency are required. From such demands, various efficient neural network paradigms, such as spiking neural networks (SNNs) or binary neural networks (BNNs), have been proposed. However, they have sticky drawbacks, such as degraded inference accuracy and latency. To solve these problems, we propose a single-step spiking neural network (S3NN), an energy-efficient neural network with low computational cost and high precision. The proposed S3NN processes the information between hidden layers by spikes as SNNs. Nevertheless, it has no temporal dimension so that there is no latency within training and inference phases as BNNs. Thus, the proposed S3NN has a lower computational cost than SNNs that require time-series processing. However, S3NN cannot adopt naïve backpropagation algorithms due to the non-differentiability nature of spikes. We deduce a suitable neuron model by reducing the surrogate gradient for multi-time step SNNs to a single-time step. We experimentally demonstrated that the obtained surrogate gradient allows S3NN to be trained appropriately. We also showed that the proposed S3NN could achieve comparable accuracy to full-precision networks while being highly energy-efficient.

Highly accurate multi-person pose estimation at a high framerate is a fundamental problem in autonomous driving. Solving the problem could aid in preventing pedestrian-car accidents. The present study tackles this problem by proposing a new model composed of a feature pyramid and an original head to a general backbone. The original head is built using lightweight CNNs and directly estimates multi-person pose coordinates. This configuration avoids the complex post-processing and two-stage estimation adopted by other models and allows for a lightweight model. Our model can be trained end-to-end and performed in real-time on a resource-limited platform (low-cost edge device) during inference. Experimental results using the COCO and CrowdPose datasets showed that our model can achieve a higher framerate (approx. 20 frames/sec with NVIDIA Jetson AGX Xavier) than other state-of-the-art models while maintaining sufficient accuracy for practical use.

Contact imaging is a compact lensless microscopy technique that allows observation of living cells placed directly on a charge-coupled device sensor. However, captured images exhibit low visibility in general. This work is aimed at integrating computational refocusing and high-frequency illumination into contact imaging to enhance the visibility of the captured images. Refocusing contributes to the synthesis of shallow depth-of-field images from the light field captured by moving a point light source. The visibility may be enhanced via high-frequency illumination. Naïve integration of both techniques requires several observations. Therefore, we propose an efficient integration from reduced observations by reformulating the computational refocusing applied in contact imaging. Furthermore, we developed a prototype system that demonstrates the effectiveness of the proposed method.

In this paper, we present a feature learning method for long-time sensor data. Although feature learning methods have been successfully used in many applications, they cannot extract features efficiently when the dimension of training data is quite large. To address this problem, we propose a method to search effective features from long-time sensor data. The important characteristic of our method is that it searches the features based on the gradient of input vectors to minimize the objective function of the learning algorithm. We apply our method to the estimation of physical capacity from wearable sensor data. The experimental results show that our method can estimate leg muscle strength more accurately than conventional methods using a feature learning method and current clinical index.

In this article, we propose a transfer learning method for deep neural networks (DNNs). Deep learning has been widely used in many applications. However, applying deep learning is problematic when a large amount of training data are not available. One of the conventional methods for solving this problem is transfer learning for DNNs. In the field of image recognition, state-of-the-art transfer learning methods for DNNs re-use parameters trained on source domain data except for the output layer. However, this method may result in poor classification performance when the amount of target domain data is significantly small. To address this problem, we propose a method called All-Transfer Deep Learning, which enables the transfer of all parameters of a DNN. With this method, we can compute the relationship between the source and target labels by the source domain knowledge. We applied our method to actual two-dimensional electrophoresis image (TDEI) classification for determining if an individual suffers from sepsis; the first attempt to apply a classification approach to TDEIs for proteomics, which has attracted considerable attention as an extension beyond genomics. The results suggest that our proposed method outperforms conventional transfer learning methods for DNNs.

Recently, methods for the unsupervised learning of features from large data sets have been attracting much attention. These methods have been especially successful in the area of computer vision. However, there is a problem that it is difficult to determine what kind of features will result in a high classification performance. Indeed, the difficulty of determining the learning parameters is a widely known problem in the field of feature learning. To address this problem, this paper presents a feature-learning method which uses classification results to progressively learn multiple features of varied complexity. The proposed method enables the learning of both simple robust features and complex features which represents difficult patterns. In addition, we assign regularization weights that are based on the complexity of the features. This modification emphasizes simple representation and prevents over fitting. Experimental results with medical image classification show that the proposed method is superior to the conventional method, especially when classification is difficult.

In this paper, we propose a new non-rigid robust registration method that registers a point distribution model (PDM) of a surface to given 3D images. The contributions of the paper are (1) a new hierarchical statistical shape model (SSM) of the surface that has better generalization ability is introduced, (2) the registration algorithm of the hierarchical SSM that can estimate the marginal posterior distribution of the surface location is proposed, and (3) the registration performance is improved by (3-1) robustly registering each local shape of the surface with the sparsity regularization and by (3-2) referring to the appearance between the neighboring model points in the likelihood computation. The SSM of a liver was constructed from a set of clinical CT images, and the performance of the proposed method was evaluated. Experimental results demonstrated that the proposed method outperformed some existing methods that use non-hierarchical SSMs.

We show how to incorporate a statistical shape model into the nonrigid ICP framework, and propose a robust nonrigid ICP algorithm. In the nonrigid ICP framework, a template surface is represented by a set of points, and the shape of the template is parametrized by a transformation matrix per one template point. In the proposed method, the statistics of the matrices are estimated based on a set of training surfaces, and the statistical shape model is incorporated into the nonrigid ICP framework by modifying the representation of the stiffness of the template. The statistical shape model and a noise model make it possible to discriminate outliers from inliers in given targets. Our proposed method detects the outliers, which are not represented by the models appropriately, based on their sparseness. The detected outliers are automatically excluded from the target to be registered, and the template is deformed to fit the inliers only. As the result, the accuracy of the registration is improved. The performance of the proposed method is evaluated qualitatively and quantitatively using synthetic data and clinical CT images.

The purpose of our study was to develop a completely automated method for detecting abdominal tumors in given FDG-PET/CT images. The purpose of this article is to report on a classifier that distinguishes malignant tumors from false positives. For constructing the classifier, we employed the kernel Fisher discriminant analysis (KFDA), and experimentally evaluated the relationship between the dimensionality of training data, the leave-one-out (LOOCV) error of the classifier, and the stability of the classifier. The results showed that, as the dimensionality increased, the LOOCV error decreased but the classifier became unstable. Our method firstly binarizes a given PET image to extract candidates of malignant tumors. From a set of the images of the extracted candidates, we constructed the classifier. For computing the classification function, we normalize the size of the extracted images of the candidates and incorporate them into training data. When we change the size of the normalized images, the dimensionality of the training data changes and we obtain different classification function. For each classification function obtained at different dimensionality, we approximately evaluated its uniform stability and computed the upper bound of the difference between the LOOCV error of the classifier and the generalization error. In this article, we show that it is needed to evaluate the stability of the classifier to determine the size of the normalized images.