This paper develops a novel methodology for using symbolic knowledge in deep learning. From first principles, we derive semantic loss function that bridges between neural output vectors and logical constraints. captures how close the network is to satisfying constraints on its output. An experimental evaluation shows it effectively guides learner achieve (near-)state-of-the-art results semi-supervised multi-class classification. Moreover, significantly increases ability of predict structured...

Few-shot learning (FSL) aims to learn new categories with a few visual samples per class. class representations are often biased due data scarcity. To mitigate this issue, we propose generate based on semantic embeddings using conditional variational autoencoder (CVAE) model. We train CVAE model base classes and use it features for novel classes. More importantly, guide VAE strictly representative by removing non-representative from the training set when show that scheme enhances...

Federated learning (FL) is a privacy-preserving distributed paradigm that enables clients to jointly train global model. In real-world FL implementations, client data could have label noise, and different vastly noise levels. Although there exist methods in centralized for tackling such do not perform well on heterogeneous settings, due the typically smaller sizes of datasets privacy requirements FL. this paper, we propose FedCorr, general multi-stage framework tackle FL, without making any...

Class-agnostic object counting aims to count instances of an arbitrary class at test time. Current methods for this challenging problem require human-annotated exemplars as inputs, which are often unavailable novel categories, especially autonomous systems. Thus, we propose zero-shot (ZSC), a new setting where only the name is available during Such system does not human annotators in loop and can operate automatically. Starting from name, method that accurately identify optimal patches then...

Data augmentation is an intuitive step towards solving the problem of few-shot classification. However, ensuring both discriminability and diversity in augmented samples challenging. To address this, we propose a feature disentanglement framework that allows us to augment features with randomly sampled intra-class variations while preserving their class-discriminative features. Specifically, disentangle representation into two components: one represents variance other encodes information. We...

Abstract Motivation Antimicrobial peptides (AMPs) have the potential to inhibit multiple types of pathogens and heal infections. Computational strategies can assist in characterizing novel AMPs from proteome or collections synthetic sequences discovering their functional abilities toward different microbial targets without intensive labor. Results Here, we present a deep learning-based method for computer-aided AMP discovery that utilizes transformer neural network architecture with...

Unsupervised domain adaptation (UDA) is vital for alleviating the workload of labeling 3D point cloud data and mitigating absence labels when facing a newly defined domain. Various methods utilizing images to enhance performance cross-domain segmentation have recently emerged. However, pseudo labels, which are generated from models trained on source provide additional supervised signals unseen domain, inadequate utilized due their inherent noisiness consequently restrict accuracy neural...

The detection of pathogens in food is an essential part quality control and safety plan. A laser light scattering system designed for rapid label-free bacterial pathogens. sample prepared with unknown contamination placed inside the where a beam focused on it. When microbial particles pass through beam, absorbed, refracted scattered by these particles. intensity measured assembly twelve sensors features were extracted using power spectrums characteristics from time domain signal. Different...

Learning unsigned distance fields (UDF) directly from raw point clouds as the implicit representation for surface reconstruction is a promising learning-based method reconstructing open surfaces and supervision-free attributes. In most UDF methods, Chamfer Distance (CD), commonly used metric in 3D domains, reckoned preferable loss function training neural networks that predict UDFs. However, CD intrinsically suffers deficiencies like insensitivity to density distribution inclination be...

3D Gaussian Splatting (3D-GS) has made a notable advancement in the field of neural rendering, scene reconstruction, and novel view synthesis. Nevertheless, 3D-GS encounters main challenge when it comes to accurately representing physical reflections, especially case total reflection semi-reflection that are commonly found real-world scenes. This limitation causes reflections be mistakenly treated as independent elements with presence, leading imprecise reconstructions. Herein, tackle this...

Diffusion models excel at high-quality image and video generation. However, a major drawback is their high latency. A simple yet powerful way to speed them up by merging similar tokens for faster computation, though this can result in some quality loss. In paper, we demonstrate that preserving important during significantly improves sample quality. Notably, the importance of each token be reliably determined using classifier-free guidance magnitude, as measure strongly correlated with...

Text-to-image diffusion models have demonstrated remarkable capability in generating realistic images from arbitrary text prompts. However, they often produce inconsistent results for compositional prompts such as "two dogs" or "a penguin on the right of a bowl". Understanding these inconsistencies is crucial reliable image generation. In this paper, we highlight significant role initial noise inconsistencies, where certain patterns are more than others. Our analyses reveal that different...

For classification tasks, deep neural networks are prone to overfitting in the presence of label noise. Although existing methods able alleviate this problem at low noise levels, they encounter significant performance reduction high or even medium levels when is asymmetric. To train classifiers that universally robust all and not sensitive any variation model, we propose a distillation-based framework incorporates new subcategory Positive-Unlabeled learning. In particular, shall assume small...

Federated edge learning enables intelligence services to be deployed at the of future wireless network. To address limited spectral resource and constrained scalability, analog over-the-air federated is recently proposed achieve fast aggregations with enhanced efficiency, privacy concurrent reduced access latency, via exploiting superposition property waveforms. From a practical aspect real-world data, noisy label quality commonly exists in local data diverse clients heterogeneous network,...

Class-agnostic counting (CAC) has numerous potential applications across various domains. The goal is to count objects of an arbitrary category during testing, based on only a few annotated exemplars. In this paper, we point out that the task interest when there are multiple object classes in image (namely, multi-class counting) particularly challenging for current models. They often greedily every regardless To address issue, propose localizing area containing via exemplar-based...

Class-agnostic object counting aims to count instances of an arbitrary class at test time. It is challenging but also enables many potential applications. Current methods require human-annotated exemplars as inputs which are often unavailable for novel categories, especially autonomous systems. Thus, we propose zero-shot (ZSC), a new setting where only the name available during This obviates need human annotators and automated operation. To perform ZSC, finding few crops from input image use...

Energy saving plays an important role in designing AI-native 6G networks. Radio Access Network (RAN) slicing is a fundamental tool to save energy through resource multiplexing. However, as the AI services required by users become more heterogenous than ever network, service-oriented RAN naturally consumes lot of energy, leading tradeoff between QoS guarantees and for network scheduler decide. In this paper, we propose sustainable (SSO) networks jointly optimize workload distribution...