This article explains the GANomaly model for semi‑supervised anomaly detection, detailing its generator‑encoder‑discriminator architecture, loss functions, testing phase scoring, and provides annotated PyTorch source code to help readers implement and understand the approach.
This article explains why traditional GANs suffer from training instability, introduces the Wasserstein (Earth‑Mover) distance as a smoother alternative, derives the WGAN objective, discusses Lipschitz constraints, and provides practical PyTorch code modifications to convert a vanilla GAN into a stable WGAN.
Ant Group’s Ant Technology Research Institute had four papers from its Visual Intelligence Lab accepted at NeurIPS 2022, covering rank diminishing in deep networks, geometry‑aware 3D image synthesis, dynamic discriminators for GANs, and uncertainty‑aware hierarchical refinement for incremental classification, highlighting the institute’s cutting‑edge AI research.
This talk presents a personalized news recommendation system that leverages knowledge graphs to enhance recommendation accuracy, explainability, and user interest modeling, detailing background, graph construction methods, multi‑task deep learning architecture, experimental results, and future research directions.
This article explains the background, workflow, and technical details of using the YOLOv5 one‑stage object detection algorithm to enable fast, accurate video analytics for urban management, covering data augmentation, backbone design, FPN‑PAN neck, and prediction output processing.
This article reviews the evolution of deep‑learning image‑classification networks, surveys attention mechanisms for fine‑grained tasks, describes the CVPR 2022 FGVC9 competition solution using RegNetY and random attention cropping, and discusses its deployment in automotive series recognition along with future challenges.
AI-powered porn detection leverages deep neural networks to classify images, but faces serious hurdles such as visual similarity with benign content, subjective standards of obscenity, and vulnerabilities stemming from training data, making human moderators indispensable for reliable content safety.
AI-based porn detection uses deep neural networks to classify images, but faces tough hurdles such as visual similarity with benign content, subjective standards for nudity, and vulnerabilities from training‑data dependence, meaning human moderators remain essential for reliable safety.
This article presents an in‑depth overview of intelligent question answering in QQ Browser search, covering its background, the core KBQA and DeepQA technologies, system architecture, challenges, recent advances such as end‑to‑end, knowledge‑guided and multimodal QA, and practical Q&A for deployment.
This article explains how AI technologies enable end‑to‑end freight safety monitoring, from pre‑trip and in‑trip risk identification to targeted interventions and governance, addressing challenges such as long‑tail data, small‑sample learning, fine‑grained classification, and multi‑level filtering.
The presentation by Tencent expert Yuan Yi details the Wuliang deep learning system for recommendation, covering its background, technical challenges such as massive data and real‑time requirements, the parameter‑server based solutions for training and inference, model compression techniques, and continuous online deployment strategies.
A visual‑lossless deep‑learning pre‑processor that employs a DCT‑based low‑rank loss and an ultra‑lightweight CPU‑friendly model achieves up to 20% bitrate reduction for 1080p videos while preserving high‑frequency details, enabling real‑time transcoding and bandwidth savings for popular content on Bilibili.
This talk explores the challenges and opportunities of leveraging sparsity in machine learning models, covering fine‑grained and coarse‑grained sparsity, NVIDIA Ampere’s 2:4 structured sparsity, knowledge‑distillation techniques for converting unstructured to structured sparsity, and model compression strategies for generative adversarial networks.
DeepMind’s AlphaFold inspired Uni‑Fold, now accelerated with Alibaba Cloud’s PAI platform, can predict protein structures up to 6.6k amino acids—covering 99.992% of known sequences—delivering ten‑minute inference for SARS‑CoV‑2 spike trimers and setting new performance benchmarks for AI‑driven structural biology.
This article presents a hybrid pipeline combining traditional computer vision techniques and deep learning models to automatically extract and verify text background colors in UI automation screenshots, effectively addressing challenges like limited training data and complex borders to significantly reduce manual inspection costs while achieving high accuracy and robustness in production environments.
iQ Dubbing, iQIYI’s AI‑driven dubbing platform, leverages deep‑learning models to automate voice‑over, international voice extraction, and sound‑effect creation, reducing dubbing turnaround from months to under a week and boosting efficiency more than threefold, as recognized by the 2022 ChinaMM Innovation Product award.
This article presents a comprehensive overview of Tencent AI Lab's graph‑based approaches for detecting misinformation and advanced persistent threats, detailing the challenges of modeling news content and social context, the design of the Post‑User Interaction Network (PSIN), experimental results on large multi‑topic datasets, and a novel graph‑pretraining pipeline for APT detection.
The Tencent Cloud‑backed “Tencent Cloud Cup” BCI competition, part of the World Robot Contest, drew over 250 teams from 26 provinces and three countries to tackle brain‑computer tasks like spelling and emotion detection, demonstrating typing, wheelchair and robotic arm applications, with the winning Chinese university team typing 81 characters in 285 seconds and results set for 5G deployment and publication in Brain Science Advances.
SepLUT, a new separable lookup‑table framework, splits color enhancement into a 1‑D LUT for independent adjustments and a 3‑D LUT for correlated changes, predicted by a lightweight CNN, enabling quantizable, real‑time ISP performance with state‑of‑the‑art results on the FiveK benchmark.
Disney’s Content Discovery team introduces M5, a multi‑modal, multi‑interest, multi‑scenario recall model that enhances VOD and live streaming recommendations by leveraging rich metadata, user behavior, and contextual features, outperforming baseline methods with significant hit‑ratio gains across Hulu and Disney+.
This article examines the evolution of mobile UI testing toward AI‑driven approaches, outlines the challenges of large‑scale apps, and details a practical workflow for constructing image‑based anomaly datasets, training a ResNet‑18 model, and iterating on detection performance.
This article introduces recurrent neural networks (RNNs) and their ability to handle sequential data, explains the limitations of vanilla RNNs, presents the LSTM architecture with its gates, and provides complete Keras code for data loading, model building, and training both vanilla RNN and LSTM models.
This article explains how the Fine‑ and Coarse‑Granularity Hybrid Self‑Attention (FCA) mechanism reduces BERT’s computational cost by over 50% while keeping accuracy loss under 1%, detailing the method, experimental results, and its significance for efficient large‑scale language models.
This article introduces artificial neural networks, explains the role of artificial neurons and their weighted connections, and provides an overview of common activation functions—including linear, nonlinear ramp, threshold/step, and sigmoid forms—highlighting their characteristics and typical saturation values.
This article provides a thorough overview of the most widely used machine learning algorithms, classifying them by learning style and problem type, and highlighting popular methods such as supervised, unsupervised, semi‑supervised, regression, instance‑based, regularization, decision‑tree, Bayesian, clustering, association rule, neural network, deep learning, dimensionality‑reduction, and ensemble techniques.
This article introduces an AI-powered makeup transfer technique that copies makeup from any reference image onto a user's photo, explains its advantages over traditional sticker-based apps, outlines a three‑step workflow, and provides sample Python code runnable on Huawei Cloud ModelArts.
This report details a series of experiments on Chinese‑English mixed‑language speech recognition, introducing language‑identification loss and language‑model integration to improve acoustic modeling, reduce mixed error rates, and achieve significant gains over a baseline end‑to‑end ASR system.
This article introduces reinforcement learning by defining agents, environments, rewards, and policies, explains key concepts such as Markov Decision Processes and Bellman equations, and surveys major algorithms—including dynamic programming, Monte‑Carlo, TD learning, policy gradients, Q‑learning, DQN, and evolution strategies—while highlighting practical challenges and notable case studies like AlphaGo Zero.
This article reviews image quality assessment fundamentals, modern CNN‑based IQA models, and their deployment at 58.com to automatically score, filter, and rank millions of recruitment photos, achieving a drop in low‑quality images from 9% to zero while boosting overall accuracy to 94.7%.
This article presents JD's comprehensive technical solution for boosting recommendation ad ranking by addressing cold‑start, shallow user interest extraction, and insufficient global data through a variational feature learning framework, enhanced user‑interest networks, and full‑domain collaborative modeling, achieving over 1% AUC gain and notable revenue growth.
The inaugural Ship‑Sea Data Intelligent Application Innovation Competition, co‑hosted by Taihu Laboratory, Huawei and Wuxi authorities, showcased cutting‑edge AI techniques—such as multi‑scale training, TTA, knowledge distillation, and model pruning—to improve surface and underwater target detection for vessels, nets, buoys, and marine life, while offering transparent rankings, research funding, and a platform for advancing maritime AI.
This article reviews the recent NLPCC 2022 NER competition, explains the evolution of named entity recognition, details the five major modeling paradigms, and describes the winning team’s relation‑classification approach, data‑augmentation strategy, experimental results, and its practical deployment in NetEase Cloud Commerce services.
This article presents an end‑to‑end AI‑driven visual testing solution for mobile applications, detailing the business pain points, data set construction, CNN‑based model design, training procedures, performance evaluation with ROC and confusion matrices, and future directions for improving defect detection accuracy.
MNN 2.0 transforms Alibaba’s lightweight deep‑learning engine into a unified edge‑cloud framework, delivering ultra‑small binaries, broad model‑format support, and aggressive CPU/GPU/DSP/NPU optimizations—including SIMD, Winograd, quantization, and sparse computation—while providing Python‑style APIs for preprocessing, inference, and on‑device training.
The article introduces Whale, an open‑source distributed training framework that unifies multiple parallelism strategies, uses hardware‑aware load balancing to accelerate giant models like BERT‑Large and the trillion‑parameter M6 on heterogeneous GPU clusters, and details its architecture, planning, and real‑world performance gains.
This presentation explains how leveraging user behavior sequences with supervised and unsupervised deep learning models, including end‑to‑end and two‑stage architectures, improves transaction fraud detection by identifying distinct patterns of account takeover and stolen‑card activities and outlines the engineering deployment pipeline.
The article details Meituan's engineering journey from small DNNs to hundred‑gigabyte deep learning models for food‑delivery ads, analyzing online latency and offline efficiency challenges and presenting distributed storage, CPU/GPU acceleration, OpenVINO, TensorRT, CodeGen, and data‑pipeline optimizations that dramatically improve throughput, memory usage, and sample‑building speed.
This article presents AMCAD, an adaptive mixed‑curvature graph embedding system that models heterogeneous e‑commerce search ad graphs in non‑Euclidean spaces, detailing its sample construction, three‑stage model architecture, offline and online experiments, and demonstrating significant performance gains over Euclidean baselines.
Bilibili’s quality‑controlled scene‑adaptive transcoding system automatically splits videos into shot‑level segments, predicts optimal encoding parameters with a deep‑learning model, applies two‑pass VMAF‑targeted encoding and ROI‑aware bitrate allocation, achieving stable visual quality, 99% accuracy, and roughly 15% bitrate reduction.
This article presents a complete image‑search solution that extracts visual features with the VGG16 model, stores them in the Milvus vector database, and provides a set of web APIs for training, querying, counting, searching, and deleting image vectors, all deployed via Docker containers.
This article presents a CVPR 2022 paper by Zhejiang University and Ant Group that introduces a label‑relation‑tree‑based Hierarchical Residual Network (HRN) for improving multi‑granularity image classification, detailing its motivation, architecture, composite loss design, extensive experiments on fine‑grained datasets, and practical impact on content‑security applications.
This article presents the JiuGe system developed by THUNLP for automatically generating Chinese classical poetry, detailing its research motivations, model architecture—including salient‑clue, working‑memory, topic‑memory, style‑transfer and reinforcement‑learning components—implementation, applications, and future directions.
This article reviews the evolution of click‑through rate (CTR) prediction models from early logistic regression and factorization machines to deep neural networks, introduces the data‑interaction based RIM (Retrieval & Interaction Machine) architecture with its search and prediction modules, and presents extensive experimental comparisons and future research directions.
This talk presents practical methods for accelerating deep model inference, detailing two case studies—text QA and speech QA—along with their technical challenges, and outlines optimization strategies such as model compression, multi‑operator fusion, matrix multiplication tuning, quantization, and dynamic batching.
Alibaba’s DaTaobao team presented six CVPR 2022 papers—covering video restoration, real‑time image enhancement, human‑object interaction detection, indirect illumination modeling, editable NeRFs, and robust NeRF reconstruction—with all code and models open‑sourced, and invites viewers to a live stream for detailed walkthroughs.
Neighbor Transformer (NFormer) introduces interactive multi‑image modeling for person re‑identification, using Landmark Agent Attention and Reciprocal Neighbor Softmax to efficiently fuse features across images, achieving state‑of‑the‑art accuracy and tighter embedding clusters on multiple benchmark datasets.
This article details the design, data pipeline, model architecture, loss functions, and evaluation metrics of a large‑scale multi‑label image classification system built for 58.com, showing how it improves semantic similarity detection, recommendation, and content moderation across diverse business domains.
This article reviews the seminal SRCNN paper, detailing its contributions, architecture, training pipeline, hyper‑parameters, and extensive experiments that show how a shallow fully‑convolutional network achieves superior PSNR and runtime compared to traditional sparse‑coding and bicubic methods.
This tutorial explains how to use the open‑source PaddleGAN Wav2Lip model to synchronise any face or avatar with arbitrary speech, covering the underlying AI principles, required installations, and step‑by‑step command‑line usage for creating high‑quality dubbing videos.
This article presents a comprehensive overview of adversarial example research applied to captcha systems, covering the definition and history of adversarial attacks, geometric‑aware generation frameworks, FGSM‑based attack variants, experimental results, trade‑offs between image quality and attack strength, and future work such as AdvGAN integration.
This article explains how a multimodal deep learning model combines static and temporal game data to predict frame rates, helping identify performance bottlenecks and improve client smoothness through feature fusion, data pipelines, and real‑time inference in modern games.
This article explains how modern video platforms combine traditional keyword retrieval with deep‑learning‑based vector retrieval, detailing model architectures, attention mechanisms, personalization features, offline experiments, and online A/B results that show significant improvements in recall, relevance, and user experience.
The paper introduces WPipe, a group‑based interleaved pipeline parallelism method that reduces memory overhead and weight‑update latency compared with PipeDream‑2BW, achieving up to 1.4× speed‑up and 36% lower memory usage while preserving model accuracy on large‑scale DNNs.
This article surveys recent deep learning approaches for matching in search and recommendation systems, presenting a unified view of matching, categorizing methods into representation learning and matching function learning, and detailing model architectures from input to output layers, while highlighting broader applications such as QA and image captioning.
This article presents an overview of graph data and graph neural networks, explains the core concepts of message‑passing GNNs, highlights DGL’s flexible API, high‑performance system design, large‑scale training capabilities and open‑source ecosystem, and outlines future plans and community resources.
The 4th International Workshop on Deep Learning Practice for High‑Dimensional Sparse and Imbalanced Data (DLP‑KDD 2022) invites submissions on deep‑learning systems, data representation, and user modeling for large‑scale sparse data, with a deadline of May 26, 2022 and acceptance notifications by June 20, 2022.
This article explains the role of activation functions in deep learning, covering their definition, why they are needed, the main categories—including linear, binary step, and various non‑linear functions such as Sigmoid, TanH, ReLU, Leaky ReLU, ELU, Softmax and Swish—along with each function's mathematical form, advantages, disadvantages, and practical usage recommendations.
From the 1940s birth of computers to today's ultra‑large pre‑training models like Baidu’s ERNIE 3.0, AI has progressed through three development waves, now driven by algorithms, compute and data, with pre‑training lowering application barriers and evolving toward larger, multimodal, and more generalizable systems.
This article reviews the evolution of recommendation models from early collaborative filtering to modern deep learning approaches, discusses core challenges such as CTR prediction, outlines user‑behavior and combination‑feature modeling techniques, introduces large‑embedding training and continuous‑feature embedding methods like AutoDis, and presents distributed training frameworks such as ScaleFreeCTR, concluding with future research directions.
The AURL framework, presented by Pu Shi, introduces alignment‑uniformity aware representation learning for zero‑shot video classification, achieving up to 28 % top‑1 accuracy gains on UCF101 and HMDB51, and has already boosted business metrics in Tencent’s advertising, search, and video‑channel recommendation systems.
EasyCV is Alibaba’s open‑source, PyTorch‑based visual modeling platform that unifies self‑supervised learning and Transformer techniques, offering a comprehensive algorithm suite, pre‑trained models, high‑performance training/inference optimizations, extensible architecture, and seamless cloud deployment for a wide range of computer‑vision tasks.
The article explains how transfer learning reduces data and time requirements in deep learning by reusing pretrained models for vision, natural language processing, and reinforcement learning, while discussing challenges such as overfitting, the need for progressive networks, entropy regularization, domain adaptation, multi‑task learning, and model distillation.
This article presents a no‑reference audio quality assessment system that leverages a bidirectional LSTM network to extract spectral features via FFT and predict perceptual scores, describing the architecture, technical advantages, data preparation, loss design, and TensorFlow implementation.
This article introduces the open‑source Python library ddddocr, explains its evolution from version 1.2.0 to 1.4.3—including OCR, target detection, and slider recognition features—and shows how it leverages deep‑learning and OpenCV to simplify captcha solving for developers.
Time series forecasting, essential across weather, finance, and commerce, involves tasks like classification, clustering, anomaly detection, and especially prediction; this article explores its definitions, evaluation metrics, traditional methods, machine‑learning approaches, deep‑learning models such as TFT, and emerging AutoML tools, offering practical insights and best practices.
This article presents the evolution of Alibaba's display advertising auction mechanisms, introducing the AIDA decision‑allocation framework, the Deep GSP multi‑objective smart auction, and the end‑to‑end Neural Auction, while discussing their economic theory, engineering platformization, business impact, and future research directions.
Kuaishou's audio‑video technology team presents a self‑developed custom video matting system that combines foreground, interactive, and video object segmentation to let creators extract arbitrary subjects without green screens, featuring adaptive cropping, multi‑stage training, and deployment across Android and iOS devices.
The article explains how Go's lock‑free sync.Pool can cut garbage‑collection overhead, shows practical FFmpeg thread‑parameter tuning that balances CPU use and latency for video filtering versus encoding, and compares PaddlePaddle's static and dynamic graph modes, including debugging tips and conversion to static.
This article reviews the limitations of traditional CNN‑based image stylization, explains how Transformer architectures overcome these issues with global context and self‑attention, and presents the novel StyTr² method with content‑aware positional encoding that achieves superior, detail‑preserving style transfer results.
The paper proposes StyTr², a Transformer‑based image style transfer method that uses content‑aware positional encoding to preserve details and improve feature representation, achieving high‑quality stylization with better content structure and style patterns.
This article introduces the AIDA framework for Alibaba's display advertising, detailing the business background, multi‑objective optimization challenges, the design of Deep GSP and Neural Auction mechanisms powered by deep learning and reinforcement learning, and outlines future technical and platform directions while also announcing recruitment opportunities.
This article documents the author’s step‑by‑step journey of building an OCR system for recognizing Chinese characters in a card‑game UI, covering game selection, technical background, data generation, deep‑learning model training with CRNN, real‑image data collection, optimization attempts, and final performance evaluation.
This article provides a thorough collection of commonly used PyTorch code snippets covering environment setup, reproducibility, GPU configuration, tensor manipulation, model building, data preprocessing, training and evaluation loops, custom loss functions, regularization techniques, learning‑rate scheduling, checkpointing, and practical tips for efficient deep‑learning development.
This article reviews the challenges of video inpainting, surveys traditional methods, and introduces the Spatial‑Temporal Transformer Network (STTN) that leverages multi‑scale attention and a Temporal Patch‑GAN discriminator, detailing its architecture, loss functions, training on Youtube‑VOS, and impressive restoration results.
The paper proposes a robust Grad‑CAM‑inspired visualization for Vision Transformers that combines attention weights and gradients to generate class‑specific saliency maps, demonstrates superior alignment with discriminative regions across ViT, Swin and Volo models, and shows a 76% false‑positive reduction in Baidu’s porn‑content risk control system.
This talk presents a deep‑learning framework that formulates financial index‑fund construction as a sparse portfolio optimization problem, solves the mixed‑integer programming via a two‑stage graph‑neural‑network pipeline, and demonstrates superior tracking performance and scalability on large‑scale index datasets.
This article explains how Kuaishou Y‑Tech combines image inpainting, semantic editing, and advanced GAN techniques to accurately locate and remove facial wrinkles—especially neck wrinkles—while preserving realistic skin texture and achieving high‑quality results suitable for production deployment.
This article describes Laiye's OCR error‑correction system, detailing the background challenges of Chinese character recognition, the analysis of three possible solutions, the chosen post‑processing approach, model architecture, training data, loss design, online inference, and experimental results showing a measurable performance boost.
The Tsinghua‑Meituan Digital Life Joint Research Institute’s academic salon will feature Associate Professor Jiang Rui presenting a cyclic generative adversarial network for probability density estimation, demonstrating how merging statistical models with deep‑learning techniques can solve core statistical problems and foster industry‑academia innovation.
This article provides a comprehensive overview of machine learning, explaining its definition, relationship to AI, workflow, historical evolution, major paradigms, real‑world applications, and the most common algorithms with practical usage tips.
JD Research Institute and the University of Sydney introduced ViTAEv2, a 600‑million‑parameter deep learning model that achieved a world‑leading 91.2% top‑1 accuracy on ImageNet Real without external data, demonstrating strong few‑shot learning, reducing labeling costs, and promising advances across many computer‑vision tasks.
This article presents JD's end‑to‑end semantic search recall pipeline, covering multi‑stage recall, a dual‑tower embedding model with multi‑head attention, a heterogeneous graph neural network (SearchGCN), a transformer‑based synonym generation system, and a joint index‑training approach that integrates product quantization to improve recall accuracy and efficiency.
This article presents the PLUG 270‑billion‑parameter Chinese language model and introduces a parameter‑efficient sparsity training (PST) framework that combines unstructured and structured pruning with low‑rank decomposition to dramatically reduce model size while preserving downstream performance.
The article provides a thorough guide to ranking in recommendation systems, detailing the pipeline architecture, sample handling challenges, extensive feature engineering categories, the evolution from collaborative filtering to deep and attention‑based models, and key optimization trade‑offs between memorization, generalization, and efficient user‑interest modeling.
This article presents a comprehensive overview of NetEase Cloud Music's advertising system, detailing its architecture, core challenges, CTR and CVR prediction models, feature engineering, model evolution from LR to deep learning, user vector modeling, and practical recommendations for improving ad performance.
The article surveys deep‑learning‑based face‑recognition systems, detailing detection, preprocessing, and recognition pipelines, describing evaluation metrics such as TAR, FAR, and Rank‑K, reviewing major datasets like LFW, MS‑Celeb‑1M and VGGFace2, and comparing leading architectures—including FaceNet, CenterLoss, SphereFace and InsightFace—while highlighting their strengths, limitations, real‑world applications, and seminal research references.
Alibaba’s MNN, a lightweight high‑performance deep‑learning inference engine, earned top honors in China’s 2022 “Science & Innovation China” awards, and delivers impressive gains such as 350% speedup on X86 CPUs, 2.1‑2.3× acceleration on ARM with sparse models, plus integrated OpenCV/Numpy functionality for edge AI deployment.
The talk presents JD's end‑to‑end semantic search recall pipeline, covering multi‑stage retrieval, a dual‑tower embedding model with multi‑head attention, a heterogeneous graph neural network for low‑frequency items, automatic synonym generation via transformer models, and a joint training approach that integrates product quantization directly into the model to improve accuracy and efficiency.
This article traces the development of embedding methods—from the early word2vec model through item2vec, DeepWalk, Node2vec, EGES, HERec, GraphRT, and target‑fitting approaches like DSSM and YouTube recommendation—highlighting how sequence‑construction and target‑fitting paradigms have shaped modern recommendation systems and AI applications.
Fluid, an open‑source CNCF project co‑developed by Alibaba Cloud and Nanjing University, introduces a dataset abstraction and elastic caching architecture that automatically optimizes I/O for cloud‑native deep‑learning training jobs, and its research was accepted as a full paper at the prestigious ICDE 2022 conference.
This article presents a curated collection of open‑source face recognition projects—including OpenFace, face_recognition, InsightFace, RetinaFace, SCRFD, FaceNet, Deepface, and CompreFace—detailing their features, GitHub stars, usage examples, and code snippets for Python and TensorFlow implementations.
This article details Alibaba's exploration of intelligent supply‑chain forecasting, covering scenario classification, three generations of prediction algorithms, the self‑developed Falcon model, performance evaluation, and real‑world cases such as Double 11 and live‑streaming, highlighting challenges and practical solutions.
This guide explains how to replace YOLOv4's CSPdarknet53 backbone with MobileNetV1, V2, or V3 networks, detailing the architecture analysis, code implementations, training setup, dataset preparation, and inference procedures for building a lightweight object detection model.
This article presents a comprehensive study of graph‑based recall techniques for content recommendation, detailing how knowledge‑graph‑augmented user‑behavior graphs and novel attention‑driven models such as GADM, SGGA, and SGGGA improve performance for users with sparse interaction histories.
The framework tags short videos by fusing text, image and audio‑video features through specialized extraction, classification, generative and retrieval modules, then ranking candidates with a multimodal BERT model, delivering accurate, business‑specific tags that boost recommendation, search and advertising.
The article surveys motion retargeting, describing its importance for virtual characters, outlining traditional geometric methods and recent deep‑learning approaches, presenting a customized Interaction Mesh‑based solution from Kuaishou Y‑tech, and discussing performance, limitations, and future research directions.
This talk presents Alibaba's travel intent prediction system, detailing the unique challenges of low‑frequency, multi‑source travel behavior, the multi‑granular CNN and time‑attention model architecture, experimental comparisons with baselines, and how integrated user interest modeling improves recommendation performance.
The article showcases three AI‑driven sports video solutions—real‑time figure‑skating action recognition with ST‑GCN, multimodal video classification merging text, image and audio via ERNIE and TextCNN, and automated football highlight clipping using TSN‑BMN‑LSTM—each achieving over 85% accuracy, fully open‑source on PaddlePaddle with one‑click notebooks and a live developer session.
This article provides a comprehensive overview of Horovod, Uber's open-source distributed deep learning framework, covering its architecture, communication mechanisms, performance benchmarks, and deployment on Kubernetes and Spark for accelerated multi-GPU training.
This article presents a comprehensive overview of 58's local service recommendation system, detailing the characteristics of its recommendation scenarios, the evolution of tag and post recommendation pipelines, the underlying deep‑learning models such as Bi‑LSTM, ATRank, DeepFM and ESMM, and outlines future research directions.
