This tutorial explains how to install, launch, and use TensorBoard and Visdom with PyTorch, providing step‑by‑step commands, code examples for logging training metrics, and visualizing images and plots to monitor deep‑learning experiments.
iQIYI’s OCR journey, explained by researcher Harlon, covers the evolution from separate detection and recognition pipelines to end‑to‑end models, key algorithms like CTPN, DB and CRNN, large‑scale simulated training, diverse video‑text applications, and future goals such as mobile deployment and tighter NLP integration.
Researchers at Northwestern University have developed the Tri‑AI system, a deep‑learning facial recognition platform that accurately identifies individual golden snub‑nosed monkeys in the wild, achieving 94% precision and enabling non‑invasive monitoring, data collection, and broader applications across multiple animal species.
The paper introduces UMEC, a unified model compression framework that jointly optimizes feature embedding and prediction modules under resource constraints, achieving up to three‑fold compression of recommendation models without sacrificing accuracy, and demonstrates superior performance on multiple benchmark datasets.
Amap’s visual‑technology pipeline automatically generates and updates POI names by crowdsourcing street‑level images, applying deep‑learning scene‑text recognition, dual‑branch classification of text attributes, and a BERT‑plus‑graph‑attention model that selects and orders recognized text, achieving about 95 % naming accuracy.
This article presents a comprehensive overview of Taobao Live's deep ranking system evolution, detailing the DBMTL multi‑task learning framework, the two‑tower DMR matching‑ranking architecture, and the RUI Ranking refer‑item model, together with their offline formulas, online deployment scenarios, and measured performance gains across click‑through, watch‑time, and conversion metrics.
This article reviews recent research on medical sentiment analysis, covering interpretability of neural models, contextual aspect‑sentiment interactions, strategies for handling noisy labels, and methods for building domain‑specific sentiment lexicons, highlighting challenges and proposed solutions.
This article explains the architecture and ranking process of modern recommendation systems, covering the two-stage pipeline of candidate generation and ranking, the evolution from rule‑based methods to logistic regression, GBDT, wide‑and‑deep, and deep learning models, and discusses challenges such as feature non‑linearity, multi‑objective optimization, and the need for post‑ranking interventions.
This article summarizes a AAAI 2021 paper that introduces a deep meta‑learning framework with an amortization network to generate spatio‑temporal representations, enabling accurate retail sales predictions across regions and time periods, especially during high‑volume shopping festivals.
PyTorch 1.8, released by the PyTorch team, bundles over 3,000 commits since 1.7, introducing AMD ROCm support, enhanced Python function conversion, stable FFT and linear‑algebra APIs, complex‑tensor autograd, distributed‑training improvements, new mobile tutorials, performance tools, and several prototype features.
This article examines why accelerating cloud‑based deep learning is essential, presents benchmark results comparing GPU generations and distributed training, introduces Alluxio as a distributed memory‑level cache, details its architecture on Kubernetes, and offers concrete tuning strategies to overcome I/O bottlenecks and boost training performance.
WeChat’s 8.0 update introduced an OCR pipeline that first quickly detects text in images, classifies the image type, applies a lightweight multi‑language detection network and a MobileNetV3‑based DBNet recognizer with a multi‑task CTC/Attention model, then merges results via a rule‑based layout analyzer to deliver accurate, well‑formatted extracted text across diverse languages and document types.
This article presents the design and implementation of a multi‑objective optimization framework using Multi‑gate Mixture‑of‑Experts (MMoE) to improve click‑through, conversion, and purchase behaviors in Taobao's "Lying Flat" home‑goods recommendation channel, detailing model variants, feature engineering, loss weighting, and online A/B test results.
This article introduces the concepts, tasks, and recent advances in text sentiment analysis, focusing on attribute‑level sentiment (TG‑ABSA) and opinion‑quadruple extraction, describing unsupervised, reading‑comprehension, and multi‑task deep‑learning approaches, their implementation on Huawei Cloud, experimental results, and future research directions.
This article details the AI-driven intelligent voice quality inspection system built by 58.com, covering its background, multi‑layer architecture, speech recognition, role and tag identification, backend services, and the resulting efficiency gains for large‑scale call‑center operations.
This article details Kuaishou Y‑tech’s Gorgeous beauty platform, covering traditional smoothing, advanced skin‑tone effects, AI‑driven blemish removal, clarity enhancement, local facial tuning, and the UNet‑based GorgeousGAN that delivers one‑click high‑definition beauty for live‑stream and short‑video applications.
This article explains Alibaba's Deep Tree Matching (TDM) technology, covering the challenges of large‑scale match retrieval, the progression from classic two‑stage recall to tree‑based indexing, max‑heap tree modeling, beam‑search retrieval, and the joint model‑index learning across TDM 1.0, 2.0, and 3.0, highlighting significant offline and online performance gains and future research directions.
This article details how 58.com tackled the challenges of multi‑business recommendation on its home page by developing a dual‑channel deep interest model, introducing customized feature‑crossing, optimizing training and online performance, and exploring multi‑channel extensions for broader scenario adaptation.
This article describes how 58.com’s rental platform leverages deep learning models such as Wide&Deep, DeepFM, DCN, DIN, and DIEN to improve search ranking, detailing data pipelines, feature engineering, model iteration, multi‑task training, prediction optimizations, and resulting online performance gains.
This guide walks you through installing NVIDIA drivers, CUDA, and nvidia-docker2 on a Linux host and configuring Docker containers to access the GPU, including verification steps and sample TensorFlow and PyTorch commands.
The article introduces the CA‑TCN model, which combines cross‑session item graphs, a temporal convolutional network, and a session‑context graph to capture both item‑level and session‑level cross‑session influences, achieving state‑of‑the‑art performance on benchmark session‑based recommendation datasets.
The lecture presented by Assistant Professor Yanzhi Wang introduces a compression‑compilation co‑design framework (CoCoPIE) that achieves real‑time deep‑learning inference on edge devices through novel pruning and quantization techniques, delivering up to 180× speedup without accuracy loss.
The AMAP‑TECH competition challenged participants to infer real‑time road conditions from in‑vehicle video, prompting the authors to combine lane‑wise vehicle detection with LightGBM and later an end‑to‑end DenseNet‑GRU model, augment data, ensemble five networks, and achieve a 0.7237 F1 score while outlining future deployment and research directions.
This article introduces a hierarchical graph neural network model that jointly captures multi‑scale temporal patterns and global spatial context for urban traffic flow prediction, demonstrates its superiority over existing methods on multiple public datasets, and validates each component through extensive ablation studies.
This article reviews the development of Zhihu's search system, describing the transition from early GBDT ranking to deep neural networks, the introduction of multi‑objective and position‑bias‑aware learning‑to‑rank methods, context‑aware techniques, end‑to‑end training, personalization, and future research directions.
This article presents an overview of graph neural networks, explains their computational framework, analyzes space and time complexities, and proposes ten practical strategies—including edge avoidance, dimensionality reduction, selective iteration, memory baking, distillation, partitioning, sparse computation, routing, and cross-sample feature sharing—to significantly lower the cost of large‑scale GNN processing.
The recall stage, the first filtering step after candidate generation, transforms intuitive attribute‑based shortcuts into sophisticated matrix‑factorization and embedding methods—such as dual‑tower and tree‑based models—enabling fast, personalized, diverse candidate selection for real‑time recommendation pipelines.
This article explains what OCR is, introduces the EasyOCR Python library, shows how to install it, walks through step‑by‑step usage with code examples, and summarizes the underlying deep‑learning techniques powering the library.
This article introduces the concept of user preference mining, discusses challenges such as accurate expression, interpretability, and high-dimensional preferences, reviews statistical and model-based approaches including weighting, decay, XGBoost, DNN, LSTM, Seq4Rec, and Deep Interest Network, and describes their practical implementation at Beike.
This article reviews four mainstream model compression and acceleration methods—structural optimization, pruning, quantization, and knowledge distillation—explaining their principles, implementations, and performance, and presents practical examples such as DistillBERT, TinyBERT, and FastBERT with comparative results.
The article presents Alibaba's display‑advertising team’s three‑stage computing‑power efficiency evolution, introduces the DCAF personalized power‑allocation algorithm with its Lagrangian formulation, and describes the AllSpark dynamic‑control framework that together enable a flexible, resource‑aware Transformers engine achieving significant business gains during high‑traffic events.
The article traces L4 autonomous-driving perception from early rule-based point-cloud methods through data-driven deep-learning models to emerging self-learning, multi-task systems, and highlights four key hurdles—model generalization and explainability, robust multi-sensor fusion, real-time compute limits, and proper uncertainty handling—calling for integrated AI, engineering, and data solutions.
The paper proposes a deep meta‑learning framework that generates spatio‑temporal representations for retail sales forecasting, especially during large shopping festivals, by combining amortization networks, shared statistical structures, and alternating spatial‑temporal training to achieve robust and accurate predictions despite scarce historical data.
tinygrad, an open‑source autograd tensor library by George Hotz, offers a compact PyTorch‑like experience in fewer than 1000 lines, with easy installation, GPU support via PyOpenCL, full EfficientNet inference, and extensible optimizers for rapid neural‑network prototyping.
Learn how to generate cinematic ghosting effects in videos by leveraging PaddlePaddle’s PaddleHub deep learning library and the pretrained deeplabv3p_xception65 model for semantic segmentation, with step‑by‑step code, environment setup, and practical testing on classic martial‑arts footage.
The article presents a comprehensive study of re‑ranking in large‑scale industrial recommendation pipelines, identifies four key challenges—context awareness, permutation specificity, computational complexity, and business constraints—and proposes a two‑stage List Retrieval System that combines fast sequence search and a generative re‑ranking network with a deep context‑wise model, achieving significant online gains across multiple Taobao feed scenarios.
This article introduces LightSeq, an open‑source, GPU‑accelerated inference engine that dramatically speeds up Transformer‑based models such as BERT and GPT by up to 14× over TensorFlow, supports multiple decoding strategies, integrates seamlessly with major deep‑learning frameworks, and provides detailed performance benchmarks and technical optimizations.
Tencent’s open‑source Game AI SDK provides a versatile automation testing platform—supporting a wide range of game genres and mobile/PC apps—by integrating environment simulation, configurable tools, image‑recognition modules, and deep‑learning algorithms (DQN and IM) into a unified, user‑friendly workflow for training and executing AI agents.
The Cross‑Session Aware Temporal Convolutional Network (CA‑TCN) combines a cross‑session item graph, a dilated temporal convolutional network, and a session‑context graph to capture both global cross‑session signals and positional order, achieving state‑of‑the‑art recommendation performance on benchmarks and slated for deployment in Meituan’s e‑commerce platforms.
This article explores how AI is being applied to waste classification—from smart trash cans and autonomous garbage trucks to deep‑learning models—while highlighting data‑labeling hurdles, model selection pitfalls, and how platforms like Huawei Cloud ModelArts can streamline development.
This article introduces the AI‑powered "Bringing-Old-Photos-Back-to-Life" project, explains its requirements, provides step‑by‑step commands for full‑pipeline restoration, scratch detection, global restoration, and face enhancement, and shares the Colab demo and GitHub repository for hands‑on experimentation.
This article provides a comprehensive overview of Generative Adversarial Networks, covering their original formulation, training dynamics, loss functions, major variants such as DCGAN and WGAN, and practical implementations for image‑to‑image translation, style transfer, and handwriting font synthesis at Laiye Technology.
Y‑Tech’s in‑house keypoint detection system powers Kuaishou’s beauty and effect filters across live streaming, video creation, and editing by leveraging lightweight deep‑learning models, extensive multi‑scenario data collection, and specialized handling of occlusion, enabling real‑time, robust facial and body landmark tracking on diverse mobile devices.
This article reviews common image similarity detection techniques—including hash-based methods (aHash, pHash, dHash), histogram comparison, feature matching with ORB and SIFT/SURF, bag‑of‑words with K‑Means, and CNN‑based VGG16 features—detailing their algorithms, Python implementations, performance characteristics, and practical considerations.
This article presents a deep‑learning Seq2Seq model that converts images of mathematical formulas—including matrices, equations, fractions, and radicals—into LaTeX sequences with over 95% accuracy, detailing data preparation, LaTeX normalization, model architecture, training, inference, and post‑processing techniques.
This article introduces NVIDIA's Merlin recommendation system framework, detailing its three core components—NVTabular for feature engineering, HugeCTR for high‑performance CTR model training, and Triton for inference—while discussing common pipeline challenges, performance advantages, and example implementations for deep‑learning‑based recommender models.
iQIYI’s AI‑driven coloring engine uses a GAN‑based model with attention and semantic reference modules to automate line‑art coloring, delivering consistent styles across frames, cutting coloring time to 0.7 seconds per image and reducing labor costs by about 30% in animation pipelines.
This article explains the fundamentals of seq2seq neural machine translation models, covering encoder‑decoder architecture, word embeddings, context vectors, RNN processing, and the attention mechanism introduced by Bahdanau and Luong, with visual illustrations and reference links for deeper study.
Dewu Photo Search evolved from a limited Aliyun‑based prototype to a self‑developed pipeline using EfficientNet detection and 128‑dim embeddings, boosting top‑1 shoe accuracy over 100 % and overall precision by up to 41 %, while reducing latency and improving scalability despite remaining stability challenges.
The presentation by Alibaba Entertainment's senior algorithm expert details the challenges of video search in the 4G/5G era and describes a comprehensive framework covering business overview, relevance and ranking, multimodal retrieval, deep semantic modeling, dataset construction, and practical deployment techniques.
This lecture details the end‑to‑end AI architecture for unmanned stores, covering algorithm module selection, calibration, face recognition, multi‑task detection, tracking, recommendation, data collection, augmentation, model training, and GPU‑accelerated deployment to achieve real‑time performance and high accuracy.
This article explains how a cloud‑based trade fair leverages AI techniques—including user and item profiling, multi‑level caching with Caffeine and Redis, and a Deep Interest Network model with attention mechanisms—to deliver personalized, high‑performance recommendations for exhibitors, buyers, and individual users.
The article surveys recent advances in single‑channel speech separation and target‑speaker extraction, explains the encoder‑separator‑decoder framework, compares frequency‑ and time‑domain methods, highlights models such as SpEx+, DPRNN‑Spe, and introduces Iterative Refined Adaptation, which iteratively improves speaker embeddings to boost SI‑SDR performance and enables effective speaker‑suppression for applications like in‑vehicle voice interaction.
This article investigates how to dynamically identify road closures by classifying vehicle, bicycle, and pedestrian trajectories, addressing sample imbalance and noisy labels with a label‑probability mixture Bayesian model and a deep‑learning image‑encoding pipeline, and compares their experimental results.
This lecture outlines the key challenges of deep learning in retail—growing model size, speed, and energy consumption—and presents a comprehensive acceleration framework covering algorithmic optimizations like network design, pruning, and hardware acceleration, with practical examples such as MobileNet, model compression, and edge deployment.
This article discusses a deep spatial-temporal tensor factorization framework for large-scale user visits understanding and forecasting, addressing challenges in advertising inventory prediction and demonstrating significant improvements over traditional methods.
This paper proposes a novel spatio‑temporal Siamese network that identifies whether GPS trajectories belong to the same driver, addressing data‑cost and scalability issues of existing methods and demonstrating superior accuracy and generalization on large‑scale taxi driver datasets.
FrameX is an AI-powered platform developed by Beike’s Data Intelligence Center that leverages vector floorplan data to automatically tag, score, interpret, cluster, and retrieve housing layouts, supporting numerous business scenarios through a layered architecture of data, feature, and application layers.
FinBERT 1.0 is an open‑source Chinese BERT model pre‑trained on large‑scale financial corpora that achieves 2‑5 % F1 improvements across multiple downstream fintech tasks without additional tuning, demonstrating the value of domain‑specific pre‑training for natural language processing.
This article presents DeepMatch, a behavior‑sequence based deep learning recall model enhanced with play‑rate and intent‑type embeddings, describes its self‑attention architecture, factorized embedding parameterization, multitask loss design, distributed TensorFlow training tricks, and demonstrates significant offline and online improvements in music recommendation performance.
The article reviews deep‑model compression methods—ADMM‑based structured pruning, low‑bit quantization, and teacher‑student knowledge distillation—and their automated AutoCompress workflow, demonstrating how these techniques shrink neural networks enough to run real‑time driver‑monitoring and other intelligent cockpit functions on resource‑limited automotive hardware while preserving accuracy.
This article traces the evolution of artificial neural networks from their biological inspiration, explains the McCulloch‑Pitts neuron model, details perceptron architecture and learning rule with a Scikit‑Learn example, and introduces multilayer perceptrons and the back‑propagation algorithm together with common activation functions.
The paper proposes classifying vehicle, bicycle, and pedestrian trajectories to detect road closures, introducing a probability‑mixture Bayesian model that mitigates label noise and class imbalance through joint feature densities, and a dual‑stream deep‑learning approach encoding trajectories as images, with experiments showing the Bayesian method outperforming the neural network on limited labeled data.
Didi’s COVID‑19 mask‑detection system, built on a DFS‑based face detector and an attention‑enhanced ResNet‑50 mask classifier achieving over 99.5 % accuracy, has been deployed in vehicles, open‑sourced, and complemented by top‑ranked results in international visual AI contests, including first place in driver‑gaze prediction and podium finishes in emotion recognition and model‑compression challenges.
This lecture details the evolution of product recognition algorithms from traditional image classification to deep‑learning‑based object detection, discusses challenges in dense retail scenes, presents solutions like rotated bounding boxes and multi‑source sensor fusion, and explains practical deployment in digital and unmanned stores.
This article reviews Kuashou Y‑Tech’s ECCV‑2020 paper on monocular depth estimation, detailing its novel GCB‑SAB network, new HC‑Depth dataset, specialized loss functions and edge‑aware training, and demonstrates superior performance on NYUv2, TUM and real‑world mobile AR applications.
This article presents a comprehensive study of using reinforcement learning to improve e‑commerce product ranking, covering the limitations of traditional scoring, the design of context‑aware models, a pointer‑network based sequence generator, various RL algorithms, extensive offline evaluations, and successful online deployment with future research directions.
This guide explains five key PyTorch tensor functions—expand, permute, tolist, narrow, and where—detailing their purpose, usage, and code examples so you can manipulate tensors efficiently on CPU or GPU.
This article introduces the Deep Sparse Network (NON), a new deep neural architecture for recommendation systems that combines field‑wise networks, across‑field interaction networks, and an operation‑fusion network, and demonstrates its superior performance through extensive experiments and ablation studies.
Didi’s Navigation system combines a novel MJO 3D panoramic map with advanced data‑compression and octree rendering, precise main/sub‑road yaw detection using LSTM‑based models trained on GPS and image data, and a lightweight deep‑learning engine optimized for mobile CPUs/GPUs, delivering accurate, real‑time guidance for ride‑hailing and autonomous driving.
The article reviews NLP interpretability techniques, contrasting white‑box approaches that probe model internals such as neuron analysis, diagnostic classifiers, and attention with black‑box strategies like rationales, adversarial testing, and local surrogates, and argues that black‑box methods are generally more practical for business deployment despite offering shallower insights.
This article explains how Baidu App leverages a VDSR‑based deep‑learning model and a series of mobile‑side optimizations to deliver real‑time image and video super‑resolution on iOS and Android devices, detailing the technical challenges, performance gains, and integration steps.
This article presents a comprehensive study of tree-based deep models for large-scale matching, introduces the theoretical framework of optimal tree models, proposes a Beam Search aware training algorithm (BSAT/OTM) to address training-test mismatch, and demonstrates significant recall improvements on Amazon Books and UserBehavior datasets.
This article presents a comprehensive study of the Wide & Deep architecture applied to user purchase‑intent quantification in the real‑estate domain, detailing feature engineering, model design, training procedures, experimental results, and extensions with GRU‑based sequential modeling to improve accuracy.
This article presents a comprehensive English summary of the deep learning specialization infographics originally shared by Andrew Ng, covering fundamentals, logistic regression, shallow and deep neural networks, regularization, optimization, hyperparameters, convolutional and recurrent networks, and practical advice for model building and evaluation.
This article describes how 58.com applied deep‑learning techniques—including feature engineering, sample construction, model evolution from Wide&Deep to DIN/DIEN and multi‑task learning—and system‑level optimizations to improve CTR/CPM performance in its large‑scale commercial ranking platform.
This article presents CSCNN, a category‑specific convolutional neural network that integrates visual priors into click‑through‑rate (CTR) models for JD.com’s e‑commerce advertising, detailing its motivation, architecture, engineering optimizations, offline and online training strategies, and empirical performance gains on both public and industrial datasets.
In this interview, Peng Changping, JD's recommendation advertising algorithm lead, explains how large‑scale machine learning and deep learning power JD's e‑commerce recommendation system, detailing its impact on revenue, key technical components, challenges, and future directions.
Learn how to create a browser‑compatible MNIST image classification model using TensorFlow.js, covering data preprocessing with sprite images, model construction, training, and evaluation, while providing complete JavaScript code examples and practical tips for handling ArrayBuffer, DataView, and visualization.
This article details the end‑to‑end deep‑learning workflow for click‑through‑rate (CTR) prediction in 58.com’s commercial ranking system, covering system architecture, feature engineering, sample construction, model evolution from Wide&Deep to DIN/DIEN, and engineering optimizations that together yielded significant CPM and CVR improvements.
This article presents Alibaba's comprehensive approach to user modeling for search CTR/CVR ranking, detailing the abstraction of user information, multi‑scale behavior processing, enhanced transformer‑based model structures, client‑side click and exposure modeling, and experimental results showing significant AUC improvements.
This article describes Suning's AI‑driven end‑to‑end solution for massive time‑series monitoring, anomaly detection, forecasting with DeepAR, MQ‑RNN, MQ‑CNN, ensemble methods, root‑cause localization using Hotspot and Monte‑Carlo Tree Search, and the evolution of its large‑scale log analytics platform.
This article explains the fundamental differences between computational advertising and recommendation systems, outlines the distinct problems each field addresses, and surveys the most widely used advertising models—including traditional machine‑learning approaches, deep‑learning architectures, and hybrid solutions—providing practical insights for engineers in both domains.
This article describes how Xiaomi's AI team tackles Automatic Speech Recognition (ASR) query errors by analyzing error patterns, employing BERT, ELECTRA and a soft‑masked BERT model, generating synthetic noisy data with a fuzzy‑phoneme generator, and presenting experimental results and future research directions.
This article presents a comprehensive overview of how 58.com leverages large‑scale voice data from call‑center, private phone, and micro‑chat platforms, detailing data collection, annotation, Kaldi‑based chain model training, lattice‑free techniques, and end‑to‑end Transformer‑CTC models to improve Chinese speech recognition performance.
GoodFuture's AI team secured the UbiComp2020 championship by using a novel sensor‑driven handwriting recognition approach that bypasses traditional OCR, showcasing China's growing strength in educational AI and advanced deep‑learning models.
This article presents a comprehensive overview of Ctrip's hotel recommendation system, covering its technical architecture, data processing pipelines, various ranking and embedding models—including FM, Wide&Deep, DeepFM, and FTRL—deployment methods such as PMML and TensorFlow Serving, offline and online evaluation results, and challenges like cold‑start and diversity.
This article surveys the landscape of automatic program synthesis for logic code, covering visual programming, example‑driven generation, code‑completion models, intent inference, NL2SQL, NL2IFTTT, and advanced frameworks like TranX and Debuild, while highlighting current challenges and research directions.
The Boundary Content Graph Neural Network (BC‑GNN) introduces a bipartite‑graph framework that jointly refines start/end boundary probabilities and segment‑content confidence, enabling more precise temporal action proposals and achieving state‑of‑the‑art results on ActivityNet‑1.3 and THUMOS14.
The article introduces EdgeRec, an edge‑computing powered recommendation framework that moves user‑interest perception and ranking to the client side to overcome latency in traditional cloud‑centric recommender systems, detailing its architecture, heterogeneous behavior modeling, attention‑based reranking, and experimental gains.
The article details the design and deployment of an AI-powered intelligent voice quality inspection system at 58.com, covering its overall architecture, speech recognition, role identification, tag detection, rechecking platform, and backend infrastructure, and demonstrates its impact on call‑center efficiency and service quality.
This article introduces an open‑source Java Spring Boot project that implements a license‑plate detection and recognition system with training capabilities, detailing its features, architecture, supported plate types, software requirements, processing steps, installation guide, and reference resources.
Gaode Maps leverages advanced scene text recognition, evolving from traditional image processing to deep‑learning based detection and recognition pipelines, integrating multi‑stage models, data augmentation, and synthetic sample generation to achieve high‑accuracy, fast POI and road data automation.
The article details the Tencent Advertising Algorithm competition final, explains the chizhu team's approach of converting ad IDs into word sequences for text classification using large‑scale word2vec embeddings and a dual BiLSTM architecture, presents custom loss functions, training tricks, and shares full Python model code, achieving an overall rank of 11.
Amap uses advanced scene text recognition combining detection and recognition modules, deep learning, data synthesis, and result fusion to automate map data production, achieving state-of-the-art performance and automating the majority of POI and road updates, significantly reducing labor costs.
This article explains how Amap leverages scene text recognition to automate map data production, detailing the evolution from traditional image algorithms to deep‑learning models, the current detection and recognition framework, performance results, and future research directions for handling blur, data scarcity, and semantic understanding.
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.
