Recent stereo matching networks achieves dramatic performance by introducing epipolar line constraint to limit the matching range of dual-view. However, in complicated real-world scenarios, the feature information based on intra-epipolar line alone is too weak to facilitate stereo matching. In this paper, we present a simple but highly effective network called Alternated Attention U-shaped Transformer (AAUformer) to balance the impact of epipolar line in dual and single view respectively for excellent generalization performance. Compared to other models, our model has several main designs: 1) to better liberate the local semantic features of the single-view at pixel level, we introduce window self-attention to break the limits of intra-row self-attention and completely replace the convolutional network for denser features before cross-matching; 2) the multi-scale alternated attention backbone network was designed to extract invariant features in order to achieves the coarse-to-fine matching process for hard-to-discriminate regions. We performed a series of both comparative studies and ablation studies on several mainstream stereo matching datasets. The results demonstrate that our model achieves state-of-the-art on the Scene Flow dataset, and the fine-tuning performance is competitive on the KITTI 2015 dataset. In addition, for cross generalization experiments on synthetic and real-world datasets, our model outperforms several state-of-the-art works.
Although large language models demonstrate remarkable question-answering performances, revealing the intermediate reasoning steps that the models faithfully follow remains challenging. In this paper, we propose FAME (FAithful question answering with MontE-carlo planning) to answer questions based on faithful reasoning steps. The reasoning steps are organized as a structured entailment tree, which shows how premises are used to produce intermediate conclusions that can prove the correctness of the answer. We formulate the task as a discrete decision-making problem and solve it through the interaction of a reasoning environment and a controller. The environment is modular and contains several basic task-oriented modules, while the controller proposes actions to assemble the modules. Since the search space could be large, we introduce a Monte-Carlo planning algorithm to do a look-ahead search and select actions that will eventually lead to high-quality steps. FAME achieves state-of-the-art performance on the standard benchmark. It can produce valid and faithful reasoning steps compared with large language models with a much smaller model size.
We proposal a new approach, namely the weight pathway analysis (WPA), to study the mechanism of multilayer neural networks. The weight pathways linking neurons longitudinally from input neurons to output neurons are considered as the basic units of a neural network. We decompose a neural network into a series of subnetworks of weight pathways, and establish characteristic maps for these subnetworks. The parameters of a characteristic map can be visualized, providing a longitudinal perspective of the network and making the neural network explainable. Using WPA, we discover that a neural network stores and utilizes information in a "holographic" way, that is, the network encodes all training samples in a coherent structure. An input vector interacts with this "holographic" structure to enhance or suppress each subnetwork which working together to produce the correct activities in the output neurons to recognize the input sample. Furthermore, with WPA, we reveal fundamental learning modes of a neural network: the linear learning mode and the nonlinear learning mode. The former extracts linearly separable features while the latter extracts linearly inseparable features. It is found that hidden-layer neurons self-organize into different classes in the later stages of the learning process. It is further discovered that the key strategy to improve the performance of a neural network is to control the ratio of the two learning modes to match that of the linear and the nonlinear features, and that increasing the width or the depth of a neural network helps this ratio controlling process. This provides theoretical ground for the practice of optimizing a neural network via increasing its width or its depth. The knowledge gained with WPA enables us to understand the fundamental questions such as what to learn, how to learn, and how can learn well.
Self-supervised learning has shown very promising results for monocular depth estimation. Scene structure and local details both are significant clues for high-quality depth estimation. Recent works suffer from the lack of explicit modeling of scene structure and proper handling of details information, which leads to a performance bottleneck and blurry artefacts in predicted results. In this paper, we propose the Channel-wise Attention-based Depth Estimation Network (CADepth-Net) with two effective contributions: 1) The structure perception module employs the self-attention mechanism to capture long-range dependencies and aggregates discriminative features in channel dimensions, explicitly enhances the perception of scene structure, obtains the better scene understanding and rich feature representation. 2) The detail emphasis module re-calibrates channel-wise feature maps and selectively emphasizes the informative features, aiming to highlight crucial local details information and fuse different level features more efficiently, resulting in more precise and sharper depth prediction. Furthermore, the extensive experiments validate the effectiveness of our method and show that our model achieves the state-of-the-art results on the KITTI benchmark and Make3D datasets.
Given a segment of time series of a system at a particular set of parameter values, can one infers the global behavior of the system in its parameter space? Here we show that by using a learning machine we can achieve such a goal to a certain extent. It is found that following an appropriate training strategy that monotonously decreases the cost function, the learning machine in different training stage can mimic the system at different parameter set. Consequently, the global dynamical properties of the system is subsequently revealed, usually in the simple-to-complex order. The underlying mechanism is attributed to the training strategy, which causes the learning machine to collapse to a qualitatively equivalent system of the system behind the time series. Thus, the learning machine opens up a novel way to probe the global dynamical properties of a black-box system without artificially establish the equations of motion. The given illustrating examples include a representative model of low-dimensional nonlinear dynamical systems and a spatiotemporal model of reaction-diffusion systems.
We found that the real time reverse transcription-polymerase chain reaction (RT-PCR) detection of viral RNA from sputum or nasopharyngeal swab has a relatively low positive rate in the early stage to determine COVID-19 (named by the World Health Organization). The manifestations of computed tomography (CT) imaging of COVID-19 had their own characteristics, which are different from other types of viral pneumonia, such as Influenza-A viral pneumonia. Therefore, clinical doctors call for another early diagnostic criteria for this new type of pneumonia as soon as possible.This study aimed to establish an early screening model to distinguish COVID-19 pneumonia from Influenza-A viral pneumonia and healthy cases with pulmonary CT images using deep learning techniques. The candidate infection regions were first segmented out using a 3-dimensional deep learning model from pulmonary CT image set. These separated images were then categorized into COVID-19, Influenza-A viral pneumonia and irrelevant to infection groups, together with the corresponding confidence scores using a location-attention classification model. Finally the infection type and total confidence score of this CT case were calculated with Noisy-or Bayesian function.The experiments result of benchmark dataset showed that the overall accuracy was 86.7 % from the perspective of CT cases as a whole.The deep learning models established in this study were effective for the early screening of COVID-19 patients and demonstrated to be a promising supplementary diagnostic method for frontline clinical doctors.
The orthodox approach for understanding a dynamical system is to establish its equation of motion, by which one can unveil its dynamical behavior at a given system parameter set, and reveal how the dynamic behavior evolves as the system parameters change. Here we show that this task can be fulfilled with a learning machine in a model-free way. We find that, based only on a segmental time series of a state variable recorded at present stage, the dynamics exhibited by the learning machine at different training stages can be mapped to the dynamics of the target system along a particular path in its parameter space following an appropriate training strategy that monotonously decreases the cost. This path is important, because along which the primary dynamical properties of the target system will emerge subsequently, in the simple-to-complex order, matching closely to the evolution of a natural system. A theoretical framework is proposed to explain the underlying mechanism. This revealed function of the learning machine opens up a novel way to probe the global dynamical properties of a black-box system without the equation of motion established artificially, and as such it might have huge potential applications. As an application example, this method is applied to infer what dynamical stages a variable star has experienced and how it will evolve in future by using the light curve observed presently.
Is it possible to generally construct a dynamical system to simulate a black system without recovering the equations of motion of the latter? Here we show that this goal can be approached by a learning machine. Trained by a set of input-output responses or a segment of time series of a black system, a learning machine can be served as a copy system to mimic the dynamics of various black systems. It can not only behave as the black system at the parameter set that the training data are made, but also recur the evolution history of the black system. As a result, the learning machine provides an effective way for prediction, and enables one to probe the global dynamics of a black system. These findings have significance for practical systems whose equations of motion cannot be approached accurately. Examples of copying the dynamics of an artificial neural network, the Lorenz system, and a variable star are given. Our idea paves a possible way towards copy a living brain.
Though the deep learning is pushing the machine learning to a new stage, basic theories of machine learning are still limited. The principle of learning, the role of the a prior knowledge, the role of neuron bias, and the basis for choosing neural transfer function and cost function, etc., are still far from clear. In this paper, we present a general theoretical framework for machine learning. We classify the prior knowledge into common and problem-dependent parts, and consider that the aim of learning is to maximally incorporate them. The principle we suggested for maximizing the former is the design risk minimization principle, while the neural transfer function, the cost function, as well as pretreatment of samples, are endowed with the role for maximizing the latter. The role of the neuron bias is explained from a different angle. We develop a Monte Carlo algorithm to establish the input-output responses, and we control the input-output sensitivity of a learning machine by controlling that of individual neurons. Applications of function approaching and smoothing, pattern recognition and classification, are provided to illustrate how to train general learning machines based on our theory and algorithm. Our method may in addition induce new applications, such as the transductive inference.
The paper addresses the image fusion problem, where multiple images captured with different focus distances are to be combined into a higher quality all-in-focus image. Most current approaches for image fusion strongly rely on the unrealistic noise-free assumption used during the image acquisition, and then yield limited robustness in fusion processing. In our approach, we formulate the multi-focus image fusion problem in terms of an analysis sparse model, and simultaneously perform the restoration and fusion of multi-focus images. Based on this model, we propose an analysis operator learning, and define a novel fusion function to generate an all-in-focus image. Experimental evaluations confirm the effectiveness of the proposed fusion approach both visually and quantitatively, and show that our approach outperforms state-of-the-art fusion methods.