Modern machine learning (ML) applications are often deployed in the cloud environment to exploit computational power of clusters. However, this in-cloud computing scheme cannot satisfy demands emerging edge intelligence scenarios, including providing personalized models, protecting user privacy, adapting real-time tasks, and saving resource cost. In order conquer limitations conventional computing, there comes rise on-device learning, which makes end-to-end ML procedure totally on devices,...

The ever-growing artificial intelligence (AI) applications have greatly reshaped our world in many areas, e.g., smart home, computer vision, natural language processing, etc. Behind these are usually machine learning (ML) models with extremely large size, which require huge data sets for accurate training to mine the value contained big data. Large ML models, however, can consume tremendous computing resources achieve decent performance and thus, it is difficult train them...

Real-time video perception tasks are often challenging on resource-constrained edge devices due to the issues of accuracy drop and hardware overhead, where saving computations is key performance improvement. Existing methods either rely domain-specific neural chips or priorly searched models, which require specialized optimization according different task properties. These limitations motivate us design a general task-independent methodology, called <i>Patch Automatic Skip Scheme</i> (PASS),...

One non-negligible flaw of the convolutional neural networks (CNNs) based single image super-resolution (SISR) models is that most them are not able to restore high-resolution (HR) images containing sufficient high-frequency information. Worse still, as depth CNNs increases, training easily suffers from vanishing gradients. These problems hinder effectiveness in SISR. In this paper, we propose Dual-view Attention Networks alleviate these for Specifically, local aware (LA) and global (GA)...

Over the past few years, edge learning has achieved significant success in mobile networks. Few works have designed incentive mechanism that motivates nodes to participate learning. However, most existing only consider myopic optimization and assume all are honest, which lacks long-term sustainability final performance assurance. In this paper, we propose Chiron, an incentive-driven Byzantine-resistant based on hierarchical reinforcement (HRL). First, our goal includes both...

Recent years, garbage sorting has attracted wide attention from all aspects of society. Various methods and policies for have been put forward, which started cultivating people's active awareness sorting. With the development machine vision deep learning, applying these techniques to intelligent will be a worthy direction. A mobile robot handling based on GPS auto-drive technology was designed in this paper. The scheme elaborated three mechanical structure, hardware system, software system....

Iteration based collaborative learning (CL) paradigms, such as federated (FL) and split (SL), faces challenges in training neural models over the rapidly growing yet resource-constrained edge devices. Such devices have difficulty accommodating a full-size large model for FL or affording an excessive waiting time mandatory synchronization step SL. To deal with challenge, we propose novel CL framework which adopts tree-aggregation structure adaptive partition ensemble strategy to achieve...

As to address the impact of heterogeneity in distributed Deep Learning (DL) systems, most previous approaches focus on prioritizing contribution fast workers and reducing involvement slow workers, incurring limitations workload imbalance computation inefficiency. We reveal that grouping into communities, an abstraction proposed by us, handling parameter synchronization community level can conquer these accelerate training convergence progress. The inspiration comes from our exploration prior...

Parameter server is a popular distributed processing paradigm for operating deep learning (DL) applications. As growing number of DL models are trained via shared clusters, machines in confrontation with the heterogeneous environment, which incurs unexpected phenomenon slow task speed called straggler. Straggler addressing crucial issue applications, since stragglers significantly hamper system performance. While many techniques have been deployed to mitigate stragglers, they may not achieve...

Adolescent idiopathic scoliosis is becoming a common spinal disorder among adolescents. The traditional methods of screening are labor-intensive and can result in unnecessary referrals radiological exposure for adolescents due to their low positive predictive value. For early abnormal posture, mobile-based cost-free, accurate radiation-free system proposed this paper. We establish database with labeled 2D unclothed back images corresponding whole-spine standing posterior-anterior X-ray...

The explosive growth of video traffic on today's Internet promotes the rise Neural-enhanced Video Streaming (NeVS), which effectively improves rate-distortion trade-off by employing a cheap neural super-resolution model for quality enhancement receiver side. Missing existing work, we reveal that NeVS pipeline may suffer from practical threat, where crucial codec component (i.e., encoder compression and decoder restoration) can trigger adversarial attacks in man-in-the-middle manner to...

The advent of edge computing has made real-time intelligent video analytics feasible. Previous works, based on traditional model architecture (e.g., CNN, RNN, etc.), employ various strategies to filter out non-region-of-interest content minimize bandwidth and computation consumption but show inferior performance in adverse environments. Recently, visual foundation models transformers have shown great environments due their amazing generalization capability. However, they require a large...

This paper provides a novel parsimonious yet efficient design for zero-shot learning (ZSL), dubbed ParsNets, in which we are interested composition of on-device friendly linear networks, each with orthogonality and low-rankness properties, to achieve equivalent or better performance against deep models. Concretely, first refactor the core module ZSL, i.e., visual-semantics mapping function, into several base networks that correspond diverse components semantic space, wherein complex...

Real-time video perception tasks are often challenging over the resource-constrained edge devices due to concerns of accuracy drop and hardware overhead, where saving computations is key performance improvement. Existing methods either rely on domain-specific neural chips or priorly searched models, which require specialized optimization according different task properties. In this work, we propose a general task-independent Patch Automatic Skip Scheme (PASS), novel end-to-end learning...

This paper provides an efficient training-free painterly image harmonization (PIH) method, dubbed FreePIH, that leverages only a pre-trained diffusion model to achieve state-of-the-art results. Unlike existing methods require either training auxiliary networks or fine-tuning large backbone, both, harmonize foreground object with painterly-style background image, our FreePIH tames the denoising process as plug-in module for style transfer. Specifically, we find very last few steps of (i.e.,...