Data assimilation for nonlinear state space models (SSMs) is inherently challenging due to non-Gaussian posteriors. We propose Deep Bayesian Filtering (DBF), a novel approach to data assimilation in nonlinear SSMs. DBF introduces latent variables $h_t$ in addition to physical variables $z_t$, ensuring Gaussian posteriors by (i) constraining state transitions in the latent space to be linear and (ii) learning a Gaussian inverse observation operator $r(h_t|o_t)$. This structured posterior design enables analytical recursive computation, avoiding the accumulation of Monte Carlo sampling errors over time steps. DBF optimizes these operators and other latent SSM parameters by maximizing the evidence lower bound. Experiments demonstrate that DBF outperforms existing methods in scenarios with highly non-Gaussian posteriors.

Robots operating in the real world face significant but unavoidable issues in object localization that must be dealt with. A typical approach to address this is the addition of compliance mechanisms to hardware to absorb and compensate for some of these errors. However, for fine-grained manipulation tasks, the location and choice of appropriate compliance mechanisms are critical for success. For objects to be inserted in a target site on a flat surface, the object must first be successfully aligned with the opening of the slot, as well as correctly oriented along its central axis, before it can be inserted. We developed the Four-Axis Adaptive Finger Hand (FAAF hand) that is equipped with fingers that can passively adapt in four axes (x, y, z, yaw) enabling it to perform insertion tasks including lid fitting in the presence of significant localization errors. Furthermore, this adaptivity allows the use of simple control methods without requiring contact sensors or other devices. Our results confirm the ability of the FAAF hand on challenging insertion tasks of square and triangle-shaped pegs (or prisms) and placing of container lids in the presence of position errors in all directions and rotational error along the object’s central axis, using a simple control scheme.

Image denoising is a representative image restoration task in computer vision. Recent progress of image denoising from only noisy images has attracted much attention. Deep image prior (DIP) demonstrated successful image denoising from only a noisy image by inductive bias of convolutional neural network architectures without any pre-training. The major challenge of DIP based image denoising is that DIP would completely recover the original noisy image unless applying early stopping. For early stopping without a ground-truth clean image, we propose to monitor JPEG file size of the recovered image during optimization as a proxy metric of noise levels in the recovered image. Our experiments show that the compressed image file size works as an effective metric for early stopping.

Variational autoencoders (VAEs) are one of the deep generative models that have experienced enormous success over the past decades. However, in practice, they suffer from a problem called posterior collapse, which occurs when the posterior distribution coincides, or collapses, with the prior taking no information from the latent structure of the input data into consideration. In this work, we introduce an inverse Lipschitz neural network into the decoder and, based on this architecture, provide a new method that can control in a simple and clear manner the degree of posterior collapse for a wide range of VAE models equipped with a concrete theoretical guarantee. We also illustrate the effectiveness of our method through several numerical experiments.

A problem that plagues robotic grasping is the misalignment of the object and gripper due to difficulties in precise localization, actuation, etc. Under-actuated robotic hands with compliant mechanisms are used to adapt and compensate for these inaccuracies. However, these mechanisms come at the cost of controllability and coordination. For instance, adaptive functions that let the fingers of a two-fingered gripper adapt independently may affect the coordination necessary for grasping small objects. In this work, we develop a two-fingered robotic hand capable of grasping objects that are offset from the gripper's center, while still having the requisite coordination for grasping small objects via a novel gear-type synchronization mechanism with a magnet. This gear synchronization mechanism allows the adaptive finger's tips to be aligned enabling it to grasp objects as small as toothpicks and washers. The magnetic component allows this coordination to automatically turn off when needed, allowing for the grasping of objects that are offset/misaligned from the gripper. This equips the hand with the capability of grasping light, fragile objects (strawberries, creampuffs, etc.) to heavy frying pan lids, all while maintaining their position and posture which is vital in numerous applications that require precise positioning or careful manipulation.

Three-dimensional (3D) reconstruction for cryogenic electron microscopy (cryo-EM) often falls into an ill-posed problem owing to several uncertainties in observations, including noise. To reduce excessive degree of freedom and avoid overfitting, the structural symmetry is often used as a powerful constraint. In the case of the helix, the entire 3D structure is determined by the subunit 3D structure and two helical parameters. There is no analytical method to simultaneously obtain both of the subunit structure and helical parameters. A common approach is to employ an iterative reconstruction in which the two optimizations are performed alternately. However, iterative reconstruction does not necessarily converge when a heuristic objective function is used for each optimization step. Also, the obtained 3D reconstruction highly depends on the initial guess of the 3D structure and the helical parameters. Herein, we propose a method for estimating the 3D structure and helical parameters that also performs an iterative optimization; however, the objective function for each step is derived from a single objective function to make the algorithm convergent and less sensitive to the initial guess. Finally, we evaluated the effectiveness of the proposed method by testing it on cryo-EM images, which were challenging to reconstruct using conventional methods.

It is challenging to grasp numerous objects with varying sizes and shapes with a single robot hand. To address this, we propose a new robot hand called the "F3 hand" inspired by the complex movements of human index finger and thumb. The F3 hand attempts to realize complex human-like grasping movements by combining a parallel motion finger and a rotational motion finger with an adaptive function. In order to confirm the performance of our hand, we attached it to a mobile manipulator - the Toyota Human Support Robot (HSR) and conducted grasping experiments. In our results, we show that it is able to grasp all YCB objects (82 in total), including washers with outer diameters as small as 6.4 mm. We also built a system for intuitive operation with a 3D mouse and grasp an additional 24 objects, including small toothpicks and paper clips and large pitchers and cracker boxes. The F3 hand is able to achieve a 98% success rate in grasping even under imprecise control and positional offsets. Furthermore, owing to the finger’s adaptive function, we demonstrate characteristics of the F3 hand that facilitate the grasping of soft objects such as strawberries in a desirable posture.

Deep learning models for semantic segmentation rely on expensive, large-scale, manually annotated datasets. Labelling is a tedious process that can take hours per image. Automatically annotating video sequences by propagating sparsely labeled frames through time is a more scalable alternative. In this work, we propose a novel label propagation method, termed Warp-Refine Propagation, that combines semantic cues with geometric cues to efficiently auto-label videos. Our method learns to refine geometrically-warped labels and infuse them with learned semantic priors in a semi-supervised setting by leveraging cycle-consistency across time. We quantitatively show that our method improves label-propagation by a noteworthy margin of 13.1 mIoU on the ApolloScape dataset. Furthermore, by training with the auto-labelled frames, we achieve competitive results on three semantic-segmentation benchmarks, improving the state-of-the-art by a large margin of 1.8 and 3.61 mIoU on NYU-V2 and KITTI, while matching the current best results on Cityscapes.

Food packing industry workers typically pick a target amount of food by hand from a food tray and place them in containers. Since menus are diverse and change frequently, robots must adapt and learn to handle new foods in a short time-span. Learning to grasp a specific amount of granular food requires a large training dataset, which is challenging to collect reasonably quickly. In this study, we propose ways to reduce the necessary amount of training data by augmenting a deep neural network with models that estimate its uncertainty through self-supervised learning. To further reduce human effort, we devise a data collection system that automatically generates labels. We build on the idea that we can grasp sufficiently well if there is at least one low-uncertainty (high-confidence) grasp point among the various grasp point candidates. We evaluate the methods we propose in this work on a variety of granular foods-coffee beans, rice, oatmeal and peanuts, each of which has a different size, shape and material properties such as volumetric mass density or friction. For these foods, we show significantly improved grasp accuracy of user-specified target masses using smaller datasets by incorporating uncertainty.

In this paper we target the problem of transferring policies across multiple environments with different dynamics parameters and motor noise variations, by introducing a framework that decouples the processes of policy learning and system identification. Efficiently transferring learned policies to an unknown environment with changes in dynamics configurations in the presence of motor noise is very important for operating robots in the real world, and our work is a novel attempt in that direction. We introduce MANGA: Method Agnostic Neural-policy Generalization and Adaptation, that trains dynamics conditioned policies and efficiently learns to estimate the dynamics parameters of the environment given off-policy state-transition rollouts in the environment. Our scheme is agnostic to the type of training method used - both reinforcement learning (RL) and imitation learning (IL) strategies can be used. We demonstrate the effectiveness of our approach by experimenting with four different MuJoCo agents and comparing against previously proposed transfer baselines.

Tensor decomposition methods are widely used for model compression and fast inference in convolutional neural networks (CNNs). Although many decompositions are conceivable, only CP decomposition and a few others have been applied in practice, and no extensive comparisons have been made between available methods. Previous studies have not determined how many decompositions are available, nor which of them is optimal. In this study, we first characterize a decomposition class specific to CNNs by adopting a flexible graphical notation. The class includes such well-known CNN modules as depthwise separable convolution layers and bottleneck layers, but also previously unknown modules with nonlinear activations. We also experimentally compare the tradeoff between prediction accuracy and time/space complexity for modules found by enumerating all possible decompositions, or by using a neural architecture search. We find some nonlinear decompositions outperform existing ones.

What is the role of unlabeled data in an inference problem, when the presumed underlying distribution is adversarially perturbed? To provide a concrete answer to this question, this paper unifies two major learning frameworks: Semi-Supervised Learning (SSL) and Distributionally Robust Learning (DRL). We develop a generalization theory for our framework based on a number of novel complexity measures, such as an adversarial extension of Rademacher complexity and its semi-supervised analogue. Moreover, our analysis is able to quantify the role of unlabeled data in the generalization under a more general condition compared to the existing theoretical works in SSL. Based on our framework, we also present a hybrid of DRL and EM algorithms that has a guaranteed convergence rate. When implemented with deep neural networks, our method shows a comparable performance to those of the state-of-the-art on a number of real-world benchmark datasets.

We propose a new regularization method based on virtual adversarial loss: a new measure of local smoothness of the conditional label distribution given input. Virtual adversarial loss is defined as the robustness of the conditional label distribution around each input data point against local perturbation. Unlike adversarial training, our method defines the adversarial direction without label information and is hence applicable to semi-supervised learning. Because the directions in which we smooth the model are only "virtually" adversarial, we call our method virtual adversarial training (VAT). The computational cost of VAT is relatively low. For neural networks, the approximated gradient of virtual adversarial loss can be computed with no more than two pairs of forward- and back-propagations. In our experiments, we applied VAT to supervised and semi-supervised learning tasks on multiple benchmark datasets. With a simple enhancement of the algorithm based on the entropy minimization principle, our VAT achieves state-of-the-art performance for semi-supervised learning tasks on SVHN and CIFAR-10.

Many continuous control tasks have bounded action spaces. When policy gradient methods are applied to such tasks, out-of-bound actions need to be clipped before execution, while policies are usually optimized as if the actions are not clipped. We propose a policy gradient estimator that exploits the knowledge of actions being clipped to reduce the variance in estimation. We prove that our estimator, named clipped action policy gradient (CAPG), is unbiased and achieves lower variance than the conventional estimator that ignores action bounds. Experimental results demonstrate that CAPG generally outperforms the conventional estimator, indicating that it is a better policy gradient estimator for continuous control tasks. The source code is available at https://github.com/pfnet-research/capg.

Factorized information criterion (FIC) is a recently developed approximation technique for the marginal log-likelihood, which provides an automatic model selection framework for a few latent variable models (LVMs) with tractable inference algorithms. This paper reconsiders FIC and fills theoretical gaps of previous FIC studies. First, we reveal the core idea of FIC that allows generalization for a broader class of LVMs, including continuous LVMs, in contrast to previous FICs, which are applicable only to binary LVMs. Second, we investigate the model selection mechanism of the generalized FIC. Our analysis provides a formal justification of FIC as a model selection criterion for LVMs and also a systematic procedure for pruning redundant latent variables that have been removed heuristically in previous studies. Third, we provide an interpretation of FIC as a variational free energy and uncover previously-unknown their relationship. A demonstrative study on Bayesian principal component analysis is provided and numerical experiments support our theoretical results.

Tubular shaped networks appear not only in medical images like X-ray-, time-of-flight MRI- or CT-angiograms but also in microscopic images of neuronal networks. We present EMILOVE (Efficient Monte-carlo Image-analysis for the Location Of Vascular Entity), a novel modeling algorithm for tubular networks in biomedical images. The model is constructed using tablet shaped particles and edges connecting them. The particles encode the intrinsic information of tubular structure, including position, scale and orientation. The edges connecting the particles determine the topology of the networks. For simulated data, EMILOVE was able to accurately extract the tubular network. EMILOVE showed high performance in real data as well; it successfully modeled vascular networks in real cerebral X-ray and time-of-flight MRI angiograms. We also show some promising, preliminary results on microscopic images of neurons.

We propose a new reconstruction procedure for X-ray computed tomography (CT) based on Bayesian modeling. We utilize the knowledge that the human body is composed of only a limited number of materials whose CT values are roughly known in advance. Although the exact Bayesian inference of our model is intractable, we propose an efficient algorithm based on the variational Bayes technique. Experiments show that the proposed method performs better than the existing methods in severe situations where samples are limited or metal is inserted into the body.

We propose a framework for expanding a given image using an interpolator that is trained in advance with training data, based on sparse bayesian estimation for determining the optimal and compact support for efficient image expansion. Experiments on test data show that learned interpolators are compact yet superior to classical ones.

Image expansion by linear filtering is attractive and widely used because of its simplicity and efficiency, and many interpolation methods fall in this category. In this study, we model filtering as linear regression from low- to high-resolution color image patches, and propose a learning-based design method of image expansion filters based on sparse Bayesian estimation. Sparseness is imposed on the filter coefficients to obtain compact supports. Image expansion is formulated as the problem of finding the predictive mean of a high-resolution patch given a low-resolution patch to expand. Since an exact evaluation of the predictive distribution is difficult, variational methods are employed to derive an efficient algorithm. Experiments on test data show that good generalization performance is obtained based on sparse filters and that color modeling improves the expansion quality.

Reinforcement learning (RL) methods based on least-squares temporal difference (LSTD) have been developed recently and have shown good practical performance. However, the quality of their estimation has not been well elucidated. In this article, we discuss LSTD-based policy evaluation from the new view-point of semiparametric statistical inference. In fact, the estimator can be obtained from a particular estimating function which guarantees its convergence to the true value asymptotically, without specifying a model of the environment. Based on these observations, we 1) analyze the asymptotic variance of an LSTD-based estimator, 2) derive the optimal estimating function with the minimum asymptotic estimation variance, and 3) derive a suboptimal estimator to reduce the computational burden in obtaining the optimal estimating function.

Sound production process is often expressed by a source-filter model, which assumes that sound signals are generated by convolution of a source signal and a synthesis filter. Although the source-filter model has been widely used, simultaneous estimation of source signals and synthesis filters is difficult due to its inherent indeterminacy. To reduce the indeterminacy, we propose a state-space model that utilizes temporal continuity of pitch and loudness. From the assumption that the synthesis filter contains the timbre-like instrument-specific features while the source signal represents time-variant components such as pitch and loudness, we can estimate the parameters of the synthesis filter and use them for instrument identification. The instrument identification experiments showed comparable or higher accuracy than the existing instrument identification methods with less number of parameters. This result supports the possibility to develop a reliable estimation method of a dynamic source-filter model.

Independent component analysis (ICA) is currently the most popularly used approach to blind source separation (BSS), the problem of recovering unknown source signals when their mixtures are observed but the actual mixing process is unknown. Many ICA algorithms assume that a fixed set of source signals consistently exists in mixtures throughout the time-series to be examined. However, real-world signals often have such difficult nonstationarity that each source signal abruptly appears or disappears, thus the set of active sources dynamically changes with time. In this paper, we propose switching ICA (SwICA), which focuses on such situations. The proposed approach is based on the noisy ICA formulated as a generative model. We employ a special type of hidden Markov model (HMM) to represent such prior knowledge that the source may abruptly appear or disappear with time. The special HMM setting t hen provides an effect ofvariable selection in a dynamic way. We use the variational Bayes (VB) method to derive an effective approximation of Bayesian inference for this model. In simulation experiments using artificial and realistic source signals, the proposed method exhibited performance superior to existing methods, especially in the presence of noise. The compared methods include the natural-gradient ICA with a nonholonomic constraint, and the existing ICA method incorporating an HMM source model, which aims to deal with general nonstationarities that may exist in source signals. In addition, the proposed method could successfully recover the source signals even when the total number of true sources was overestimated or was larger than that of mixtures. We also propose a modification of the basic Markov model into a semi-Markov model, and show that the semi-Markov one is more effective for robust estimation of the source appearance.

In this letter, we propose a noisy nonlinear version of independent component analysis (ICA). Assuming that the probability density function (p. d. f.) of sources is known, a learning rule is derived based on maximum likelihood estimation (MLE). Our model involves some algorithms of noisy linear ICA (e. g., Bermond & Cardoso, 1999) or noise-free nonlinear ICA (e. g., Lee, Koehler, & Orglmeister, 1997) as special cases. Especially when the nonlinear function is linear, the learning rule derived as a generalized expectation-maximization algorithm has a similar form to the noisy ICA algorithm previously presented by Douglas, Cichocki, and Amari (1998). Moreover, our learning rule becomes identical to the standard noise-free linear ICA algorithm in the noiseless limit, while existing MLE-based noisy ICA algorithms do not rigorously include the noise-free ICA. We trained our noisy nonlinear ICA by using acoustic signals such as speech and music. The model after learning successfully simulates virtual pitch phenomena, and the existence region of virtual pitch is qualitatively similar to that observed in a psychoacoustic experiment. Although a linear transformation hypothesized in the central auditory system can account for the pitch sensation, our model suggests that the linear transformation can be acquired through learning from actual acoustic signals. Since our model includes a cepstrum analysis in a special case, it is expected to provide a useful feature extraction method that has often been given by the cepstrum analysis.