This article explains how Gaode Maps leverages traditional and deep‑learning based scene text recognition techniques—including character detection, sequence models, data synthesis, and multi‑stage frameworks—to automate POI and road data production with high precision and speed.
This article describes a practical workflow for expanding a limited industrial solar‑panel defect dataset using flip augmentation, ModelArts smart labeling, and targeted data‑balancing techniques, then evaluates the impact on a ResNet‑50 classifier with detailed accuracy and recall metrics, demonstrating how thoughtful augmentation can improve defect detection performance.
DLFlow, an end‑to‑end framework from Didi’s user‑profile team, merges Spark and TensorFlow to automate feature preprocessing, large‑scale distributed training, and massive prediction for big‑data offline tasks, offering configuration‑driven pipelines, task scheduling, and easy deployment that dramatically speeds model development.
This article details Likee's short‑video recommendation pipeline, covering the evolution of its feature‑engineering framework, the use of Flink for minute‑level statistical and second‑level session features, the integration of automatic statistical features into DNN models, multimodal feature extraction, and the cache‑tensor technique that dramatically improves online inference performance.
This article examines the challenges of deep learning inference, outlines general optimization methodologies—including system-level and model-level techniques—presents practical case studies such as Transformer translation model improvements, and discusses future trends in automated compilation and performance tuning for AI services.
This article details the end‑to‑end technical architecture of WeChat's "Look" personalized recommendation service, covering data collection, recall, multi‑stage ranking, various CTR and multi‑objective models, reinforcement‑learning based mixing, diversity optimization, and the engineering hurdles overcome to deploy these solutions at massive scale.
This article reviews the practical use of embedding technologies in Tencent's mobile news recommendation pipeline, covering the fundamentals of embeddings, their historical development, item and image embeddings, user embeddings, various vector‑based recall methods, clustering strategies, and recent advances and challenges.
This article details the technical architecture behind WeChat's "Look" content recall, covering content sourcing, profiling, multimodal tagging, knowledge‑graph representations, propensity and target detection, multi‑stage recall pipelines, and a range of deep learning models including sequence, translation, BERT, dual‑tower, hybrid, and graph neural network approaches.
This article details how Jike's recommendation team upgraded from Spark to TensorFlow, introduced a twin‑tower embedding model for recall, deployed it with TensorFlow Serving and Elasticsearch, and achieved a 33.75% lift in user interaction on the dynamic square.
This article examines six core data‑science concepts—Big Data, Machine Learning, Data Mining, Deep Learning, Artificial Intelligence, and Data Science itself—explaining their definitions, interrelationships, and how they fit together as pieces of a larger analytical puzzle.
This technical talk introduces video object segmentation tasks, reviews leading datasets and state-of-the-art deep learning models, and shares practical network design rules and performance‑boosting techniques, presented by Prof. Wang Xinggang as part of Alibaba's MEDIA AI challenge series.
This article explains the fundamentals of recommendation systems, discusses early hotness‑based approaches, describes modern architectures with recall and ranking stages, reviews collaborative‑filtering techniques, matrix factorization, deep learning models such as NCF and NeuMF, and details how these methods are applied and optimized for live‑streaming services.
Paddle.js 1.0, Baidu's web‑oriented deep‑learning library, enables developers to run pretrained Paddle models directly in WebGL‑compatible browsers, offering GPU‑accelerated inference, model conversion tools, multimedia preprocessing, and a collection of ready‑made demos for on‑device AI applications.
This comprehensive review explains visual object tracking in computer vision, covering its definition, core sub‑problems of candidate generation, feature extraction, and decision making, system architecture, motion, feature and observation models, algorithm classifications, evaluation metrics, datasets, and recent research trends.
The article presents a comprehensive overview of Autohome's one‑stop machine learning platform, detailing its background, architecture, resource scheduling, data processing, model training (including distributed deep learning), deployment, real‑world applications such as purchase‑intent and recommendation models, and future development directions.
This article explains why VoIP calls suffer from brief interruptions caused by packet loss, compares traditional forward error correction (FEC) and packet loss concealment (PLC) techniques, introduces Tencent's deep‑learning‑enhanced cPLC, and shows how it significantly improves MOS scores, especially under burst‑loss conditions.
The Hybrid Spatio‑Temporal Graph Convolutional Network (H‑STGCN) integrates planned traffic flow from navigation data via a domain transformer and a compound adjacency matrix, enabling graph‑based spatio‑temporal modeling that consistently outperforms baselines in real‑world traffic forecasting and reduces severe ETA errors.
The article reviews the open‑source qa_match tool from 58.com, detailing its deep‑learning based question‑answer matching architecture, hierarchical knowledge‑base support, lightweight pre‑training model SPTM, and practical applications, while summarizing the live‑stream presentation and Q&A session.
In the fourth Alibaba Digital Media Technology Night Talk, algorithm engineer Liu Xiaolong presents an overview of video‑based temporal event detection, covering its problem background, representative prior works, and the latest research advances within the MEDIA AI Algorithm Challenge series.
This article provides a comprehensive overview of ensemble learning, explaining its principles, common classifiers, major ensemble strategies such as bagging, boosting, and stacking, and demonstrates practical deep‑learning ensemble techniques like Dropout, test‑time augmentation, and Snapshot ensembles with code examples.
This article presents a comprehensive overview of 58's recruitment recommendation system, detailing business challenges, multi‑stage recall strategies, vector‑based deep retrieval, cost‑sensitive loss design, session optimization, online incremental training, extensive offline and online evaluations, and practical lessons for future improvements.
This article details how Amap leverages deep‑learning techniques—including temporal and auxiliary feature engineering, multi‑stage RNN models, Wide&Deep architectures, and an Attention‑TCN approach—to accurately identify and handle expired points of interest, improving map freshness and user experience.
Tencent Cloud’s senior engineer Li Kaibin outlines the cloud‑based face‑effects platform, detailing its AI‑driven features such as face fusion, beauty, virtual makeup, segmentation and age‑gender transformation, the CNN‑based model training pipeline, a layered service architecture with elastic scaling and robust monitoring, and future expansions into video effects, international regions and low‑code integration.
This article explains why visualizing CNN models is crucial for understanding and debugging deep learning systems, outlines three main visualization approaches—basic architecture, activation‑based, and gradient‑based methods—and provides step‑by‑step Keras code examples, including model summary, filter visualization, occlusion mapping, saliency maps, and class activation maps.
The article introduces qa_match V1.1, an open‑source deep‑learning QA matching tool that adds one‑level knowledge‑base support, releases a lightweight Bi‑LSTM pre‑trained language model (SPTM), and provides detailed architecture, training data, performance benchmarks, future plans, and contribution guidelines.
The article explains how advertising inventory forecasting—predicting how many users will see a specific ad—poses challenges due to fluctuating traffic and user segmentation, and describes a new deep spatial‑temporal tensor factorization model that dramatically improves prediction accuracy, scalability, and robustness for large‑scale ad platforms.
Alibaba’s local life smart logistics team presents a KDD 2020 oral paper detailing a deep‑learning model for order fulfillment cycle time (OFCT) estimation in on‑demand food delivery, describing feature engineering, rider encoding, post‑processing operators, and experimental results that significantly improve prediction accuracy and user experience.
The article introduces qa_match V1.1, an open‑source, Apache‑licensed lightweight question‑answer matching system that adds a simple pre‑trained language model (SPTM), supports one‑level knowledge bases, details model architecture, training resources, performance benchmarks, future plans, and contribution guidelines.
The article presents qa_match, a TensorFlow‑based open‑source QA matching system that combines BiLSTM‑attention domain classification with DSSM intent matching, explains its architecture, training workflow, features, future roadmap, and how developers can contribute to the project.
This article reviews the evolution of semantic matching models for travel question‑answering, covering traditional keyword and probabilistic methods, deep‑learning encoders such as LSTM, CNN, and Transformer, interaction‑based architectures like MatchPyramid and hCNN, as well as transfer‑learning and multilingual extensions to improve practical deployment.
This article presents JD's e‑commerce search system, detailing the semantic vector retrieval and product ranking pipelines, the two‑tower deep learning architecture, attention‑based personalization, negative sampling strategies, training optimizations, and real‑world performance gains achieved in production.
This article describes how the 58 used‑car team built a speaker‑verification pipeline—covering data collection, MFCC feature extraction, LSTM and GMM modeling, threshold tuning, multi‑speaker clustering, and deployment results—to automatically block nuisance telemarketing calls while preserving user privacy.
This article introduces FashionBERT, a multimodal BERT‑based model that replaces ROI‑based image tokens with uniform image patches to overcome e‑commerce specific challenges, details its architecture, adaptive loss balancing, deployment in Alibaba search, and reports significant performance gains on public and internal datasets.
This article explains two model‑independent learning paradigms—multi‑task learning and transfer learning—detailing their motivations, sharing mechanisms, training procedures, theoretical formulations, and practical benefits such as improved generalization, data efficiency, and domain‑invariant representations.
The JD ARVR technology department announced two peer‑reviewed papers—one on a novel defocus blur detection network published in Transaction on Multimedia and another on a transductive relation‑propagation network for few‑shot learning accepted at IJCAI 2020—highlighting their advanced AI research and future AR‑VR ecosystem plans.
This article provides a comprehensive technical overview of modern speech recognition, covering Didi’s driver‑assistant and smart‑customer‑service applications, fundamental ASR concepts, classic GMM‑HMM methods, deep‑learning breakthroughs such as DNN‑HMM, CTC, attention‑based and transformer models, practical training tricks, signal‑processing steps, and multimodal fusion techniques.
Gaode’s automated production line combines deep‑learning image recognition, GPS‑enhanced location services, image differencing with semantic filtering, and standardized data‑fusion to continuously refresh China’s national base map, cutting manual effort and costs while delivering real‑time, high‑quality map updates for road traffic infrastructure.
Based on academic and industry experience, iQIYI has designed a deep semantic representation learning framework that integrates multimodal side information and deep models such as Transformers and graph neural networks, improving recall, ranking, deduplication, diversity and semantic matching across recommendation and search scenarios.
CenterMask is a single‑shot, anchor‑free instance segmentation framework that predicts a coarse shape from each object’s center point and a full‑image saliency map, multiplies them to produce precise masks, and achieves competitive COCO AP while running faster than two‑stage methods like Mask R-CNN.
This article examines the DeepMatch retrieval model for Tmall Genie music recommendation, detailing how incorporating user feedback such as play‑rate and query intent signals via multi‑task learning and feedback‑aware self‑attention improves recall accuracy and reduces negative recommendations, while also discussing embedding factorization, loss functions, and distributed training optimizations.
The article reviews recent research on representing massive sparse features in click‑through‑rate (CTR) models, introducing Alibaba's Res‑embedding method and Google's Neural Input Search (NIS) approach, and discusses how these techniques improve embedding efficiency and model generalization in large‑scale recommendation systems.
This tutorial demonstrates how to use Python and the PaddlePaddle deep‑learning platform to automatically remove backgrounds from multiple photos in one step, covering installation, verification, and a concise five‑line code example for batch human segmentation.
This article reviews recent innovations in CTR prediction for an intelligent marketing platform, covering deep spatio‑temporal networks, deep memory networks, and a feature‑expression‑assisted learning framework, with system architecture details, experimental results, and references to KDD and IJCAI papers.
An AI‑driven pipeline—using DAIN for frame interpolation, ESRGAN for super‑resolution, and DeOldify for colorization—transformed a low‑resolution 1920s black‑and‑white Beijing footage into a smooth, 4K, fully colored video, showcasing both technical challenges and cultural impact.
The project develops a deep‑learning pipeline for Amap’s expired POI detection that integrates two‑year temporal trend features, industry and verification attributes, a variable‑length LSTM, a Wide‑Deep architecture, and a Wide‑Attention Temporal Convolutional Network, achieving higher accuracy and efficiency while outlining future macro‑and micro‑level enhancements.
This article explains knowledge tracing and its deep learning variant DKT, detailing the underlying RNN/LSTM models, data encoding, loss functions, experimental results, practical applications in adaptive learning, as well as advantages, limitations, and future research directions.
This article introduces AnimeGAN, a lightweight GAN that converts real photos into Japanese anime‑style illustrations, explains its architecture, loss functions, model size advantages, and provides step‑by‑step instructions with code for setting up, training, and testing the TensorFlow implementation.
This article surveys the application of attention mechanisms in deep learning recommendation systems, reviewing models such as AFM, DIN, DIEN, DSIN, Behavior Sequence Transformer, Deep Spatio‑Temporal Networks, and ATRank, and discusses their architectures, attention types, advantages, and limitations.
This article presents a comprehensive overview of Youku's video search system, covering business background, evaluation metrics, system and algorithm frameworks, relevance and ranking feature engineering, dataset construction, semantic matching, multimodal video understanding, and practical case studies that illustrate the impact of deep learning and AI techniques on search performance.
This article details the evolution of Feizhu's "Guess You Like" ranking system, moving from a linear FTRL model to several end‑to‑end deep learning versions—including PALM, FB‑PALM, and GLA—highlighting technical challenges, architectural changes, and measurable performance gains.
This article presents the design, challenges, and experimental evaluation of DSTN (with pooling, self‑attention, and interactive‑attention variants) and MA‑DNN models for CTR prediction, highlighting how temporal and contextual ad information improves accuracy and yields significant online gains in large‑scale advertising systems.
In this article, Wang Zhe addresses fifteen common questions about recommendation systems, covering topics such as building cross‑domain knowledge, the role of deep reinforcement learning, handling sparse or low‑sample data, offline‑online evaluation, knowledge graphs, graph neural networks, model interpretability, large‑scale ID embedding, and career advice for engineers.
This article analyzes Huajiao Live's end‑to‑end recommendation pipeline, covering basic concepts, recall and ranking algorithms—including collaborative filtering, matrix factorization, deep learning models—and multi‑objective optimization, while detailing the engineering workflow for training, deployment, and real‑time serving in a live‑streaming environment.
DeepMatch is an open‑source Python library that implements several mainstream deep‑learning based recall‑matching algorithms, provides easy installation via pip, detailed usage examples with code, and supports exporting user and item vectors for ANN search, making it ideal for rapid experimentation and learning in recommendation systems.
The article explains why dedicated AI chips (ASICs) are needed, compares their performance and power efficiency to traditional CPUs and GPUs, describes the architecture of Google's TPU and other AI accelerators, and provides historical context for the evolution of AI hardware.
This article surveys state‑of‑the‑art deep‑learning techniques for scene text detection and recognition in e‑commerce poster images, detailing models such as CTPN, TextBoxes, SegLink, EAST, and end‑to‑end frameworks, and discusses their architectures, strengths, limitations, and future challenges.
This article describes how 58 Tongzhen leverages AI technologies—including data pipelines, feature engineering, various recall and ranking models, and AB‑testing—to build a personalized feed recommendation system for the down‑market, detailing its overall architecture, data sources, model iterations, performance gains, and future directions.
This article details the commercial strategy team's exploration of embedding technologies for a second‑hand car platform, covering mainstream embedding methods, their application in advertising recall and ranking pipelines, system architecture, model optimizations, evaluation results, and future directions.
This article explains heterogeneous computing and compares major processing units—CPU, GPU, FPGA, ASIC, and NPU—highlighting their architectures, strengths, and typical use cases, especially in deep‑learning and AI workloads.
dl_inference is an open‑source inference platform that simplifies deployment of TensorFlow and PyTorch models in production, offering unified gRPC access, load‑balanced multi‑node serving, GPU/CPU options, customizable pre‑ and post‑processing, and extensible architecture for future AI workloads.
This article explains how Huajiao Live designed and evolved its live‑streaming recommendation system, covering basic concepts, collaborative‑filtering and matrix‑factorization techniques, deep‑learning models, ranking and multi‑objective optimization, and practical deployment considerations for real‑time personalized content delivery.
This article explains the cold‑start problem in recommendation systems, outlines common industry solutions such as popular‑content, group‑representative, auxiliary‑information, bandit algorithms, and deep learning, and details how Huajiao Live applied these techniques to improve new‑user engagement and metrics.
This article details how Alibaba's 1688 B2B platform leverages deep learning techniques—including Deep Match, DIN, DIEN, DMR, and heterogeneous network models—to evolve its product recall, ranking, and live‑content recommendation pipelines, highlighting system architecture, practical lessons, and online performance improvements.
This article provides a comprehensive QA‑style review of entity‑relation extraction (ERE), covering pipeline drawbacks, various decoding strategies for NER, common relation‑classification techniques, shared‑parameter and joint‑decoding models, recent transformer‑based approaches, challenges such as overlapping entities, low‑resource settings, and the use of graph neural networks.
This article presents a comprehensive technical overview of Meitu's personalized push notification pipeline, detailing the evolution of embedding methods (Word2Vec, Airbnb listing embedding, graph embedding), multiple recall strategies (global, personalized, attribute, and content‑based), and a progression of ranking models from logistic regression to field‑wise three‑tower architectures, highlighting their impact on click‑through rates.
This article presents a comprehensive overview of Phoenix News's AI‑driven feed recommendation system, detailing its business challenges, multi‑stage architecture, deep learning models, feature pipelines, metric trade‑offs, cold‑start solutions, and practical insights for improving user satisfaction and content quality.
The article presents the Deep Match to Rank (DMR) model, which integrates collaborative‑filtering inspired user‑to‑item relevance modeling into the ranking stage of recommendation systems, achieving significant offline and online improvements in click‑through rate and revenue metrics for e‑commerce platforms.
This article describes the end‑to‑end recommendation pipeline of the Province Money Fast Report app, covering business background, data collection, model training and evaluation, the evolution from FM to DeepFM, DIN, DCN, xDeepFM, ESMM and custom networks, as well as serving strategies and practical lessons learned.
This article explains the rapid rise of artificial intelligence, outlines China's national AI development plan, compares human and machine intelligence, discusses current AI applications, future job impacts, and the infrastructure that will power the next industrial revolution.
The article introduces qa_match, an open‑source lightweight QA matching tool built on TensorFlow that combines BiLSTM‑based domain classification, DSSM‑based intent matching, and a model‑fusion strategy to deliver accurate, multi‑type responses for intelligent customer service applications.
This article presents a system‑level approach that performs deep‑learning inference directly on JPEG frequency components, uses a gating mechanism to select important DCT coefficients, and demonstrates higher accuracy with far lower bandwidth for image classification and instance segmentation tasks.
The article reviews recent deep‑learning breakthroughs in personalized content recommendation, covering news and e‑commerce systems, JD Digits' multi‑dimensional user behavior prediction models, knowledge‑graph meta‑learning, and the impact of multimodal AI on future recommendation technologies.
This paper introduces a method for video multi-label classification that incorporates label correlation features using graph convolutional networks, significantly improving classification performance.
The paper proposes the Deep Match to Rank (DMR) model, which integrates matching‑stage collaborative‑filtering ideas into the ranking stage to explicitly represent user‑to‑item relevance, thereby enhancing personalization and achieving significant CTR and DPV improvements in e‑commerce recommendation scenarios.
This article demonstrates how to build a TensorFlow‑based CNN to recognize fixed‑length 18‑digit Chinese ID card numbers, covering environment setup, synthetic data generation, model architecture, training procedure, and achieved accuracy of over 84%.
This article describes the TPNN deep‑learning platform’s multi‑GPU acceleration, data‑parallel BMUF training, LSTM‑CTC acoustic modeling, and a suite of mobile‑side optimizations—including model pruning, 8‑bit quantization, low‑precision matrix multiplication and mixed‑precision computation—that together achieve over 92% recognition accuracy for children’s English speech on both server and mobile devices.
This article presents a Bayesian smoothing approach to alleviate cold-start problems in click-through rate estimation, introduces key-value memory networks to incorporate prior knowledge, and proposes methods to convert continuous features into dictionary embeddings for deep learning models in recommendation systems.
Deep learning models often suffer from slow training and deployment due to their size, but a range of advanced acceleration methods—including model architecture optimization, pruning, quantization, knowledge distillation, and distributed training techniques—can dramatically improve speed and efficiency while maintaining performance.
A new AI algorithm jointly developed by Alibaba Security and Zhejiang University learns fine‑grained, attribute‑aware similarity embeddings for fashion images, enabling accurate detection of local design plagiarism and improving retrieval performance across multiple benchmark datasets.
The article introduces iQIYI's double‑DNN ranking architecture that combines a high‑performance teacher network with a lightweight student network through online knowledge distillation, detailing the evolution of deep learning‑based ranking models, the motivation for model upgrades, training pipelines, and experimental results that demonstrate significant latency reduction and ROI improvement.
This article reviews perception technologies used in autonomous heavy‑truck systems—including lane‑line detection, obstacle detection, and LiDAR sensing—detailing traditional and deep‑learning approaches, practical challenges on high‑speed highways, and the cost, performance, and reliability issues faced when moving these solutions to mass production.
This article reviews the rapid development of deep learning, its pivotal role in autonomous driving, outlines end‑to‑end perception‑to‑control pipelines, discusses the strengths and limitations of deep models, and proposes practical strategies such as task decomposition, multi‑method fusion, and sensor integration to improve safety and interpretability.
The article describes Ctrip's Prophet platform, which combines Flink real‑time stream processing with TensorFlow deep‑learning models to provide intelligent, low‑latency anomaly detection, replacing traditional rule‑based alerts and addressing challenges such as holiday traffic and model scalability.
During the stay‑at‑home period, this article suggests nine concise programming books covering Python neural networks, web crawling, deep learning with PyTorch, machine learning fundamentals, zero‑trust network security, classic programming pearls, Python mathematics, AI algorithms, and Vim text processing, each with brief descriptions and images.
This article presents a comprehensive tutorial on building and training a multi‑layer neural network in Python, covering data preprocessing, model architecture definition, parameter initialization, forward and backward propagation, cost computation, and parameter updates with code examples for activation functions and optimization techniques.
Meituan’s Turing platform consolidates the entire delivery‑order workflow—from massive data ingestion and feature generation to model training, evaluation, deployment, real‑time prediction, and AB testing—into a single, end‑to‑end system that evolved from a minimal MVP into a fully platformized solution, addressing speed, accuracy, and engineering‑algorithm decoupling while planning deeper deep‑learning integration.
This article presents a comprehensive overview of modern speech recognition technology, covering basic ASR concepts, classic acoustic and language models, deep‑learning approaches such as DNN‑HMM, CTC, attention‑based and transformer models, multimodal fusion, signal‑processing pipelines, and practical deployment considerations at Didi.
This article reviews the evolution of ranking models in modern recommendation systems, covering traditional linear models, factorization machines, tree‑based GBDT+LR, and a range of deep learning architectures such as Wide&Deep, DeepFM, DCN, xDeepFM, DIN, as well as multi‑task frameworks like ESMM and MMOE, and finally illustrates their practical deployment in a live streaming platform.
This article describes how a BERT‑based model is fine‑tuned to compute sentence‑pair similarity for improving recommendation accuracy in an online school, detailing the architecture, training mechanisms, code implementation, experimental results, and future extensions such as sentiment analysis.
This article provides a comprehensive overview of voice conversion technology, covering its definition, parallel and non‑parallel data approaches, classic and deep‑learning methods such as DTW, GMM, seq2seq, PPG, VAE, Flow, GAN, and practical applications and challenges in iQIYI’s products.
MedicalNet, Tencent’s open-source framework, aggregates diverse small 3D medical imaging datasets into a large pre-training corpus, applies dataset filtering and joint spatial-pixel normalization, and provides encoder-decoder models that accelerate convergence and boost accuracy for AI-driven diagnosis in data-scarce medical imaging scenarios.
This article details the evolution of ground‑marker recognition technology in high‑precision maps, covering challenges of diverse and worn markings, traditional segmentation methods, deep‑learning breakthroughs such as R‑FCN, cascade detectors, corner‑point detection, semantic segmentation, PAnet, and 3‑D point‑cloud approaches, and their impact on accuracy and production efficiency.
The 2019 CCF Big Data & Computing Intelligence Contest’s Video Copyright Detection track, judged by iQIYI, saw 705 teams from 25 countries compete, with Hengyang Data’s VGG‑16‑based solution winning, followed by Boyun Vision, Xiao Jia’s Lao Liang, Hulu Brothers and Beihang University, showcasing diverse deep‑learning and unsupervised approaches for robust video copyright detection.
Voice conversion (VC) technology changes a speaker’s timbre and style while keeping the spoken text unchanged, supporting one‑to‑one, many‑to‑one, and many‑to‑many scenarios for medical assistance and entertainment, using parallel or non‑parallel data through methods such as DTW‑aligned frame mapping, attention‑based neural networks, PPG‑LSTM pipelines, VAEs, normalizing‑flow models, and GANs, with iQIYI focusing on non‑parallel data, prosody preservation, and noise‑robust augmentation.
This article explains the concept, mathematics, and practical PyTorch implementation of dilated (or atrous) convolutions, showing how to keep image dimensions unchanged while dramatically increasing the receptive field and discussing their advantages and typical applications.
Alibaba's DAMO Academy details its AI‑driven image cutout system, describing why automated matting is needed, the four‑module pipeline (filtering, classification, detection, segmentation), architectural innovations such as dual decoders and fusion networks, and how these advances enable product‑level batch background removal.
Alibaba's DAMO Academy presents an AI‑driven image cutout system that combines filtering, classification, detection, and advanced segmentation to automate high‑precision matting, improve design efficiency, and unlock new commercial opportunities across e‑commerce and media industries.
Recent industrial advances in multi-task learning for recommendation systems, including Alibaba's ESMM and DUPN, Meituan's deep ranking, Google's MMoE, YouTube's multi-objective ranking, and Zhihu's multi-goal model, demonstrate how shared embeddings and specialized loss functions improve CTR, CVR, and user engagement metrics.
This article examines the challenges of deepening Graph Convolutional Networks (GCN), introduces ResGCN, DenseGCN, and skip‑neighbor designs to enable deeper architectures, presents experimental results showing improved performance with 28‑layer models, and outlines future research directions.
This article surveys the development of recall modules in large‑scale recommendation systems, covering traditional item‑based collaborative filtering, single‑embedding DNN and dual‑tower approaches, multi‑interest capsule networks, graph‑based embeddings, long‑short term interest modeling, and the tree‑structured TDM framework for efficient deep matching.
