Ground-truth Distribution Research Articles

Linear Cross-Entropy Certification of Quantum Computational Advantage in Gaussian Boson Sampling

Validation of quantum advantage claims in the context of Gaussian boson sampling (GBS) currently relies on providing evidence that the experimental samples genuinely follow their corresponding ground truth, i.e., the theoretical model of the experiment that includes all the possible losses that the experimenters can account for. This approach to verification has an important drawback: it is necessary to assume that the ground truth distributions are computationally hard to sample, that is, that they are sufficiently close to the distribution of the ideal, lossless experiment, for which there is evidence that sampling, either exactly or approximately, is a computationally hard task. This assumption, which cannot be easily confirmed, opens the door to classical algorithms that exploit the noise in the ground truth to efficiently simulate the experiments, thus undermining any quantum advantage claim. In this work, we argue that one can avoid this issue by validating GBS implementations using their corresponding ideal distributions directly. We explain how to use a modified version of the linear cross-entropy, a quantity that we call the LXE score, to find reference values that help us assess how close a given GBS implementation is to its corresponding ideal model. Finally, we analytically compute the score that would be obtained by a lossless GBS implementation. Published by the American Physical Society 2024

Fusing spatial and frequency features for compositional zero-shot image classification

Compositional Zero-Shot Learning (CZSL) is a particular Zero-Shot Learning (ZSL) task that aims to utilize known concepts (e.g., states and objects) to identify novel state-object compositions for Image Classification. Previous works have primarily focused on disentangling concept compositions or exploring the complex interactions between the states and objects while neglecting the critical fact that the inference of many states and compositions is related to different frequency components, which should be analyzed from the global perspective. Therefore, we propose a Spatial-frequency Feature Fusion Network (SFFNet) to introduce a new branch that utilizes a frequency-domain filtering encoder to enhance key frequency components and capture non-local interactions adaptively. Besides, we also find that the widely used backbone in conventional CZSL settings behaves superior in perceiving local features. Thus, we construct a fusion block to combine both strengths to capture the local and non-local information. In addition, the traditional one-hot ground-truth distribution in the training phase does not reflect the accurate relationships between compositions, so we propose a composition-relation based label distribution regularization to encourage the model to actively learn the inner relationships between compositions, and extend this method to construct unseen composition pseudo distribution to further enhance the model’s generalization ability to unseen compositions. Extensive experiments and detailed analysis are conducted on three popular datasets, and the results show that our method can achieve state-of-the-art performance, which reveals its superiority in identifying novel compositions. Code is available at https://github.com/lisuyi/SFFNet_czsl.

DiffuseRT: predicting likely anatomical deformations of patients undergoing radiotherapy

Objective. Predicting potential deformations of patients can improve radiotherapy treatment planning. Here, we introduce new deep-learning models that predict likely anatomical changes during radiotherapy for head and neck cancer patients. Approach. Denoising diffusion probabilistic models (DDPMs) were developed to generate fraction-specific anatomical changes based on a reference cone-beam CT (CBCT), the fraction number and the dose distribution delivered. Three distinct DDPMs were developed: (1) the image model was trained to directly generate likely future CBCTs, (2) the deformable vector field (DVF) model was trained to generate DVFs that deform a reference CBCT and (3) the hybrid model was trained similarly to the DVF model, but without relying on an external deformable registration algorithm. The models were trained on 9 patients with longitudinal CBCT images (224 CBCTs) and evaluated on 5 patients (152 CBCTs). Results. The generated images mainly exhibited random positioning shifts and small anatomical changes for early fractions. For later fractions, all models predicted weight losses in accordance with the training data. The distributions of volume and position changes of the body, esophagus, and parotids generated with the image and hybrid models were more similar to the ground truth distribution than the DVF model, evident from the lower Wasserstein distance achieved with the image (0.33) and hybrid model (0.30) compared to the DVF model (0.36). Generating several images for the same fraction did not yield the expected variability since the ground truth anatomical changes were only in 76% of the fractions within the 95% bounds predicted with the best model. Using the generated images for robust optimization of simplified proton therapy plans improved the worst-case clinical target volume V95 with 7% compared to optimizing with 3 mm set-up robustness while maintaining a similar integral dose. Significance. The newly developed DDPMs generate distributions similar to the real anatomical changes and have the potential to be used for robust anatomical optimization.

GRTR: Gradient Rebalanced Traffic Sign Recognition for Autonomous Vehicles

Traffic sign recognition is a crucial aspect of autonomous vehicle research, and deep learning techniques have significantly contributed to its progress. Nevertheless, the distribution of traffic sign information in natural complex road conditions is long-tailed, and traffic sign identification in complex road conditions has become a significant barrier to autonomous vehicle applications. The imbalanced distribution of information on the dataset migrates to the feature space during training, resulting in imbalanced classifier prediction. In this paper, we propose the gradient rebalanced traffic sign recognition (GRTR) method to address this problem for the first time. GRTR first evaluates the prediction and classification bias of the classifier using the fitted deviation between the model’s output probability and the ground-truth distributions. Then, GRTR dynamically adjusts the correction and compensation factors following the classifier’s prediction and classification biases. GRTR rebalances the positive and negative sample gradients for each category based on the synergistic effect of the correction and compensation factors to prevent the transfer of distribution imbalance and to significantly enhance the performance of the traffic sign classifier under difficult road conditions. Experimental results demonstrate that our GRTR achieves state-of-the-art performance on long-tailed traffic sign and multilabel datasets. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Note to Practitioners</i> —Most traffic sign recognition algorithms are still designed based on the assumption of a balanced distribution of traffic signs in the dataset. Real-world autonomous vehicles require traffic sign recognition on datasets with severely imbalanced distributions. This paper proposes a general approach to solving the long-tailed traffic sign recognition problem.

Digital-SMLM for precisely localizing emitters within the diffraction limit

Abstract Precisely pinpointing the positions of emitters within the diffraction limit is crucial for quantitative analysis or molecular mechanism investigation in biomedical research but has remained challenging unless exploiting single molecule localization microscopy (SMLM). Via integrating experimental spot dataset with deep learning, we develop a new approach, Digital-SMLM, to accurately predict emitter numbers and positions for sub-diffraction-limit spots with an accuracy of up to 98 % and a root mean square error as low as 14 nm. Digital-SMLM can accurately resolve two emitters at a close distance, e.g. 30 nm. Digital-SMLM outperforms Deep-STORM in predicting emitter numbers and positions for sub-diffraction-limited spots and recovering the ground truth distribution of molecules of interest. We have validated the generalization capability of Digital-SMLM using independent experimental data. Furthermore, Digital-SMLM complements SMLM by providing more accurate event number and precise emitter positions, enabling SMLM to closely approximate the natural state of high-density cellular structures.

Knowledge Distillation Based on Fitting Ground-Truth Distribution of Images

Knowledge distillation based on the features from the penultimate layer allows the student (lightweight model) to efficiently mimic the internal feature outputs of the teacher (high-capacity model). However, the training data may not conform to the ground-truth distribution of images in terms of classes and features. We propose two knowledge distillation algorithms to solve the above problem from the directions of fitting the ground-truth distribution of classes and fitting the ground-truth distribution of features, respectively. The former uses teacher labels to supervise student classification output instead of dataset labels, while the latter designs feature temperature parameters to correct teachers’ abnormal feature distribution output. We conducted knowledge distillation experiments on the ImageNet-2012 and Cifar-100 datasets using seven sets of homogeneous models and six sets of heterogeneous models. The experimental results show that our proposed algorithms improve the performance of penultimate layer feature knowledge distillation and outperform other existing knowledge distillation methods in terms of classification performance and generalization ability.

Nonbinary 2D Distribution Tool Maps Autonomic Nerve Fiber Clustering in Lumbosacral Ventral Roots of Rhesus Macaques.

Neuromodulation of the peripheral nervous system (PNS) by electrical stimulation may augment autonomic function after injury or in neurodegenerative disorders. Nerve fiber size, myelination, and distance between individual fibers and the stimulation electrode may influence response thresholds to electrical stimulation. However, information on the spatial distribution of nerve fibers within the PNS is sparse. We developed a new two-dimensional (2D) morphological mapping tool to assess spatial heterogeneity and clustering of nerve fibers. The L6-S3 ventral roots (VRs) in rhesus macaques were used as a model system to map preganglionic parasympathetic, γ-motor, and α-motor fibers. Random and ground truth distributions of nerve fiber centroids were determined for each VR by light microscopy. The proposed tool allows for nonbinary determinations of fiber heterogeneity by defining the minimum distance between nerve fibers for cluster inclusion and comparisons with random fiber distributions for each VR. There was extensive variability in the relative composition of nerve fiber types and degree of 2D fiber heterogeneity between different L6-S3 VR levels within and across different animals. There was a positive correlation between the proportion of autonomic fibers and the degree of nerve fiber clustering. Nerve fiber cluster heterogeneity between VRs may contribute to varied functional outcomes from neuromodulation.

Semi-Supervised Adversarial Learning for Attribute-Aware Photo Aesthetic Assessment

Aesthetic attributes are crucial for aesthetics because they explicitly present some photo quality cues that a human expert might use to evaluate a photo's aesthetic quality. However, annotating aesthetic attributes is a time-consuming, costly, and error-prone task, which leads to the issue that photos available are partially annotated with attributes. To alleviate this issue, we propose a novel semi-supervised adversarial learning method for photo aesthetic assessment from partially attribute-annotated photos, which can greatly reduce the reliance on manual attribute annotation. Specifically, the proposed method consists of a score-attributes generator R, a photo generator G, and a discriminator D. The score-attributes generator learns the aesthetic score and attributes simultaneously to capture their dependencies and construct better feature representations. The photo generator reconstructs the photo by feeding aesthetic attributes, score, and informative feature representation. A discriminator is used to force the convergence of the features-attributes-score tuples generated from the score-attributes generator, the photo generator, and the ground-truth distribution in labeled data for training data. The proposed method significantly outperforms the state of the art, increasing the Spearman rank-order correlation coefficient (SRCC) from the existing best reported of 0.726 to 0.761 on Aesthetic and attributes database and 0.756 to 0.774 on Aesthetic visual analysis database, respectively.

Instance-specific Semantic Augmentation for Long-tailed Image Classification.

Recent long-tailed classification methods generally adopt the two-stage pipeline and focus on learning the classifier to tackle the imbalanced data in the second stage via re-sampling or re-weighting, but the classifier is easily prone to overconfidence in head classes. Data augmentation is a natural way to tackle this issue. Existing augmentation methods either perform low-level transformations or apply the same semantic transformation for all instances. However, meaningful augmentations for different instances should be different. In this paper, we propose feature-level augmentation (FLA) and pixel-level augmentation (PLA) learning methods for long-tailed image classification. In the first stage, the feature space is learned from the original imbalanced data. In the second stage, we model the semantic within-class transformation range for each instance by a specific Gaussian distribution and design a semantic transformation generator (STG) to predict the distribution from the instance itself. We train STG by constructing ground-truth distributions for instances of head classes in the feature space. In the third stage, for FLA, we generate instance-specific transformations by STG to obtain feature augmentations of tail classes for fine-tuning the classifier. For PLA, we use STG to guide pixel-level augmentations for fine-tuning the backbone. The proposed augmentation strategy can be combined with different existing long-tail classification methods. Extensive experiments on five imbalanced datasets show the effectiveness of our method.

Deep Learning and Registration-Based Mapping for Analyzing the Distribution of Nodal Metastases in Head and Neck Cancer Cohorts: Informing Optimal Radiotherapy Target Volume Design.

We introduce a deep-learning- and a registration-based method for automatically analyzing the spatial distribution of nodal metastases (LNs) in head and neck (H/N) cancer cohorts to inform radiotherapy (RT) target volume design. The two methods are evaluated in a cohort of 193 H/N patients/planning CTs with a total of 449 LNs. In the deep learning method, a previously developed nnU-Net 3D/2D ensemble model is used to autosegment 20 H/N levels, with each LN subsequently being algorithmically assigned to the closest-level autosegmentation. In the nonrigid-registration-based mapping method, LNs are mapped into a calculated template CT representing the cohort-average patient anatomy, and kernel density estimation is employed to estimate the underlying average 3D-LN probability distribution allowing for analysis and visualization without prespecified level definitions. Multireader assessment by three radio-oncologists with majority voting was used to evaluate the deep learning method and obtain the ground-truth distribution. For the mapping technique, the proportion of LNs predicted by the 3D probability distribution for each level was calculated and compared to the deep learning and ground-truth distributions. As determined by a multireader review with majority voting, the deep learning method correctly categorized all 449 LNs to their respective levels. Level 2 showed the highest LN involvement (59.0%). The level involvement predicted by the mapping technique was consistent with the ground-truth distribution (p for difference 0.915). Application of the proposed methods to multicenter cohorts with selected H/N tumor subtypes for informing optimal RT target volume design is promising.

Frame-Recurrent Video Crowd Counting

Since video data contains temporal information, video crowd counting demonstrates more potential than single-frame crowd counting for scenarios requiring high accuracy. However, learning robust relationships among frames efficiently and cheaply is very challenging. Existing methods for video crowd counting lack explicit temporal correlation modeling and robustness, and they are complex. In this paper, we propose the Frame-Recurrent Video Crowd Counting (FRVCC) framework to solve these issues. Specifically, we design a frame-recurrent manner to recursively relate the density maps in the temporal dimension, which efficiently explores long-term inter-frame knowledge and ensures the continuity of feature map responses. FRVCC consists of three plug-in modules: an optical flow estimation module, a single-frame counting module, and a density map fusion module. For the fusion module, we propose the ResTrans network to robustly learn complementary features between visual-based and correlation-based feature maps through residual strategy and vision transformer. To constrain the output distribution to be consistent with the ground truth distribution, we introduce an adversarial loss to rectify the training process. Additionally, we release a large-scale synthetic video crowd-counting dataset, CrowdXV, to evaluate the proposed method and further improve its performance. We have conducted extensive experiments on several video-counting datasets. The results demonstrate that FRVCC achieves state-of-the-art performance and, concurrently, high generalization, high flexibility, and less complexity.

A dataset of multi-source and multi-temporal remote sensing data of cash crop planting structure in Yangling Agricultural Demonstration Zone

Satellite remote sensing technology can obtain the distribution of ground objects on a large scale in a timely manner, and provide great data and technical support for the acquisition of information on the planting structure of cash crops. Taking Yangling Agricultural Demonstration Area as the research area, this dataset is composed of four parts: remote sensing data, ground truth data, Yangling boundary and classification results. The remote sensing data consist of satellite data, such as Sentinel-2, Gaofen-1 (including Gaofen-1C satellite), Gaofen-2, and Gaofen-6 from April to September in 2021 after radiation correction, atmospheric correction, and remote sensing image processing such as orthorectification, image fusion, and image registration. Through on-the-spot investigation, visual interpretation of Google Earth, and near-ground remote sensing of UAVs in small areas, we established the ground truth distribution verification area. In terms of quality control, the remote sensing data are characteristic of little overall cloud content, uniform color, and a spatial resolution of 2m; the ground truth map, authentic and reliable, is drawn through field surveys. The dataset has been verified by random forest algorithm, and the overall classification accuracy is 86.17%. It can provide training samples for the research and application of related algorithms in the acquisition of cash crop planting structure, and can also provide data support for land use classification and changes as well as crop growth monitoring in Yangling Demonstration Zone.

Demystifying Randomly Initialized Networks for Evaluating Generative Models

Evaluation of generative models is mostly based on the comparison between the estimated distribution and the ground truth distribution in a certain feature space. To embed samples into informative features, previous works often use convolutional neural networks optimized for classification, which is criticized by recent studies. Therefore, various feature spaces have been explored to discover alternatives. Among them, a surprising approach is to use a randomly initialized neural network for feature embedding. However, the fundamental basis to employ the random features has not been sufficiently justified. In this paper, we rigorously investigate the feature space of models with random weights in comparison to that of trained models. Furthermore, we provide an empirical evidence to choose networks for random features to obtain consistent and reliable results. Our results indicate that the features from random networks can evaluate generative models well similarly to those from trained networks, and furthermore, the two types of features can be used together in a complementary way.

Feature Distribution Fitting with Direction-Driven Weighting for Few-Shot Images Classification

Few-shot learning has received increasing attention and witnessed significant advances in recent years. However, most of the few-shot learning methods focus on the optimization of training process, and the learning of metric and sample generating networks. They ignore the importance of learning the ground-truth feature distributions of few-shot classes. This paper proposes a direction-driven weighting method to make the feature distributions of few-shot classes precisely fit the ground-truth distributions. The learned feature distributions can generate an unlimited number of training samples for the few-shot classes to avoid overfitting. Specifically, the proposed method consists of two optimization strategies. The direction-driven strategy is for capturing more complete direction information that can describe the feature distributions. The similarity-weighting strategy is proposed to estimate the impact of different classes in the fitting procedure and assign corresponding weights. Our method outperforms the current state-of-the-art performance by an average of 3% for 1-shot on standard few-shot learning benchmarks like miniImageNet, CIFAR-FS, and CUB. The excellent performance and compelling visualization show that our method can more accurately estimate the ground-truth distributions.

Interior Individual Trajectory Simulation with Population Distribution Constraint

Individual trajectory generation plays an important role in simulation tasks, reconstructing fine-grained mobility behaviors that can be used to evaluate epidemic risks, congestion risks, or commercial profit. Previous research works adopt the Newton’s mechanic-based particle model as their core algorithm, such as the Social Force model. However, real-world human mobility behaviors hardly follow the particle models, especially in the interior scenes where interactions between pedestrians and environments matter. In this article, we propose a Social Force-based trajectory simulator for interior scenarios that improve both trajectory quality and generation speed for interior scenarios. First, we introduce prior scene knowledge to guide the generation process, where pedestrians are armed with exploration behaviors that follow the group-level distribution. It provides more flexibility to simulate complicated human behaviors rather than straight-line movements, generating high-quality individual trajectories. Experiments show that the correlation between the aggregated population distribution of generated trajectories and ground-truth distribution is improved by 11.84% by our method. Second, we optimize the algorithm procedure by introducing a caching mechanism for tenderized intermediate values, along with graph-processing-unit-based implementation. Compared with the baseline Social Force model, we reduced the time consumption by 95%. More importantly, based on our simulation paradigm, we quantitatively evaluate several common mobility interventions in our simulation scenario, which can shed light on better policy designs in public spaces.

Capturing the few-shot class distribution: Transductive distribution optimization

Few-shot learning is challenging since only a few labeled samples are available for training a learning model. To alleviate the data limitation problem in few-shot learning, several works try to generate samples or features by learning a model or distribution. But complex models and biased estimation of class distribution hamper their interpretability and generalization ability, respectively. In this work, we propose a generation based Transductive Distribution Optimization (TDO) method, which introduces neither extra parameters nor complex models. We use a few labeled samples and some high-confident unlabeled samples of the target set to capture the distributions of the few-shot classes, and then generate sufficient samples from them to augment the labeled inputs. Our method can work with most pre-trained feature extractors and outperforms state-of-the-art methods with a simple linear classifier. The visualization of the generated samples shows that our method can capture an accurate distribution even though the few labeled samples deviate from the ground-truth distribution.

Normative models for neuroimaging markers: Impact of model selection, sample size and evaluation criteria

Modelling population reference curves or normative modelling is increasingly used with the advent of large neuroimaging studies. In this paper we assess the performance of fitting methods from the perspective of clinical applications and investigate the influence of the sample size. Further, we evaluate linear and non-linear models for percentile curve estimation and highlight how the bias-variance trade-off manifests in typical neuroimaging data.We created plausible ground truth distributions of hippocampal volumes in the age range of 45 to 80 years, as an example application. Based on these distributions we repeatedly simulated samples for sizes between 50 and 50,000 data points, and for each simulated sample we fitted a range of normative models. We compared the fitted models and their variability across repetitions to the ground truth, with specific focus on the outer percentiles (1st, 5th, 10th) as these are the most clinically relevant.Our results quantify the expected decreasing trend in variance of the volume estimates with increasing sample size. However, bias in the volume estimates only decreases a modest amount, without much improvement at large sample sizes. The uncertainty of model performance is substantial for what would often be considered large samples in a neuroimaging context and rises dramatically at the ends of the age range, where fewer data points exist. Flexible models perform better across sample sizes, especially for non-linear ground truth.Surprisingly large samples of several thousand data points are needed to accurately capture outlying percentiles across the age range for applications in research and clinical settings. Performance evaluation methods should assess both bias and variance. Furthermore, caution is needed when attempting to go near the ends of the age range captured by the source data set and, as is a well known general principle, extrapolation beyond the age range should always be avoided. To help with such evaluations of normative models we have made our code available to guide researchers developing or utilising normative models.

Rethinking Collaborative Metric Learning: Toward an Efficient Alternative Without Negative Sampling.

The recently proposed Collaborative Metric Learning (CML) paradigm has aroused wide interest in the area of recommendation systems (RS) owing to its simplicity and effectiveness. Typically, the existing literature of CML depends largely on the negative sampling strategy to alleviate the time-consuming burden of pairwise computation. However, in this work, by taking a theoretical analysis, we find that negative sampling would lead to a biased estimation of the generalization error. Specifically, we show that the sampling-based CML would introduce a bias term in the generalization bound, which is quantified by the per-user Total Variance (TV) between the distribution induced by negative sampling and the ground truth distribution. This suggests that optimizing the sampling-based CML loss function does not ensure a small generalization error even with sufficiently large training data. Moreover, we show that the bias term will vanish without the negative sampling strategy. Motivated by this, we propose an efficient alternative without negative sampling for CML named Sampling-Free Collaborative Metric Learning (SFCML), to get rid of the sampling bias in a practical sense. Finally, comprehensive experiments over seven benchmark datasets speak to the supriority of the proposed algorithm.

Prototype-Based Self-Adaptive Distribution Calibration for Few-Shot Image Classification

Deep learning has flourished in large-scale supervised tasks. However, in many practical conditions, rich and available labeled data are a luxury. Thus, few-shot learning (FSL) has recently received boosting interest and achieved significant progress, which can learn new classes from several labeled samples. The advanced distribution calibration approach estimates the ground-truth distribution of few-shot classes by reusing the statistics of auxiliary data. However, there is still a significant discrepancy between the estimated distributions and ground-truth distributions, and artificially set hyperparameters cannot be adapted to different application scenarios (i.e., datasets). This paper proposes a prototype-based self-adaptive distribution calibration framework for estimating ground-truth distribution accurately and self-adaptive hyperparameter optimization for different application scenarios. Specifically, the proposed method is divided into two components. The prototype-based representative mechanism is for obtaining and utilizing more global information about few-shot classes and improving classification performance. The self-adaptive hyperparameter optimization algorithm searches robust hyperparameters for the distribution calibration of different application scenarios. The ablation studies verify the effectiveness of the various components of the proposed framework. Enormous experiments are conducted on three standard benchmarks such as miniImageNet, CUB-200-2011, and CIFAR-FS. The competitive results and compelling visualizations indicate that the proposed framework achieves state-of-the-art performance.

EdgeAI: Machine learning via direct attached accelerator for streaming data processing at high shot rate x-ray free-electron lasers

We present a case for low batch-size inference with the potential for adaptive training of a lean encoder model. We do so in the context of a paradigmatic example of machine learning as applied in data acquisition at high data velocity scientific user facilities such as the Linac Coherent Light Source-II x-ray Free-Electron Laser. We discuss how a low-latency inference model operating at the data acquisition edge can capitalize on the naturally stochastic nature of such sources. We simulate the method of attosecond angular streaking to produce representative results whereby simulated input data reproduce high-resolution ground truth probability distributions. By minimizing the mean-squared error between the decoded output of the latent representation and the ground truth distributions, we ensure that the encoding layers and resulting latent representation maintains full fidelity for any downstream task, be it classification or regression. We present throughput results for data-parallel inference of various batch sizes, some with throughput exceeding 100 k images per second. We also show in situ training below 10 s per epoch for the full encoder–decoder model as would be relevant for streaming and adaptive real-time data production at our nation’s scientific light sources.