This article explains the background of GPU communication, introduces DMA and RDMA fundamentals, describes how GPUDirect RDMA enables direct GPU-to-GPU memory access across machines, and presents performance results showing reduced latency and increased bandwidth for distributed deep‑learning training.
JD AI Research Institute introduced the Improved Selective Refinement Network (ISRN), a face‑detection model that outperforms all competitors on the WIDER FACE benchmark across Easy, Medium, and Hard subsets with AP scores of 96.3%, 95.4% and 90.3%, thanks to novel training strategies, COCO pre‑training, redesigned input modules, and multi‑stage classification‑regression enhancements.
This article explains the background of GPU communication, details NVIDIA's GPUDirect and its Peer‑to‑Peer features, discusses virtualization challenges, and presents performance measurements on an Alibaba Cloud GN5 instance showing latency reduction and near‑linear scaling for deep‑learning workloads.
This article explains how Alibaba's automatic task‑pipeline system transforms deep‑learning inference for retail video streams by decoupling model execution from scheduling, using Python‑based pipelines, high‑performance shared memory, and robust fault‑tolerance, achieving up to 13% faster processing and double the camera capacity.
The article traces AI’s origins from the 1956 Dartmouth conference, explains how deep learning has become the driving force behind modern AI applications, discusses the need for a mindset shift, lifelong learning, and strategic adaptation to stay relevant in the rapidly evolving intelligent era.
Cheng Huige’s winning AI Challenger 2018 solution treated fine‑grained Chinese review sentiment as a 20‑aspect multi‑class task, combining a high‑capacity LSTM encoder with self‑attention, word‑and‑character embeddings, simplified ELMo pre‑training, diverse tokenizations and a weighted seven‑model ensemble (including BERT), which together delivered the competition’s top F1 performance.
Tencent’s industrial practice shows how a large‑scale offline‑nearline‑online “Shield” recommendation architecture, powered by the DeepR framework built on RCaffe, uses deep semantic embeddings, massive neural networks and reinforcement‑learning decisions to handle billions of daily requests, demonstrating that data richness and engineering capability, not model depth alone, drive performance in big‑data recommendation systems.
The Meituan‑Dianping search team progressed its core ranking from linear, FM and GBDT models to a knowledge‑graph‑enhanced deep‑learning architecture, culminating in the listwise LambdaDNN network that directly optimizes NDCG, supported by extensive feature engineering, distributed TensorFlow training, and the Athena diagnostic system.
This article presents a detailed case study of building a purchase‑user prediction model by integrating Convolutional Neural Networks for feature extraction with Long Short‑Term Memory networks for time‑series forecasting, covering background, model structure, data augmentation, experimental results, and business impact.
The talk explains how Meituan transformed its O2O hotel search by layering a multi‑stage retrieval pipeline with intent‑aware NLP, then progressively upgrading ranking—from XGBoost to MLPs, feature‑embedding networks, and finally a Wide‑Deep multi‑task model—while tackling data sparsity, diverse scenarios, and deploying the system via TensorFlow‑Serving and the in‑house MLX platform.
This article provides a comprehensive, step‑by‑step explanation of the Transformer architecture, covering its encoder‑decoder structure, self‑attention, multi‑head attention, positional encoding, residual connections, and training processes, illustrated with diagrams and code snippets to aid readers new to neural machine translation.
This article introduces core deep‑learning architectures—including MLP, RNN, CNN, auto‑encoders, and RBM—explains common activation and loss functions, and then surveys fourteen influential deep‑learning‑based recommendation algorithms such as FM, wide&deep, deepFM, NCF, GBDT+LR, seq2seq and YouTube DNN, complete with model diagrams and reference links.
iQIYI built a cloud‑based deep‑learning training platform called Jarvis, replacing the initial Runonce service, by containerizing GPU tasks, adopting Ceph S3 storage with FUSE, optimizing data pipelines, and addressing compute, storage, and networking challenges to improve scalability and reduce GPU idle time.
iQIYI’s NLP pipeline—combining CRF‑based segmentation, bidirectional LSTM/GRU models with attention and a CNN classifier—automatically extracts opinion targets, sentiment words and polarity from unstructured video comments, aggregates them across users to reveal collective attitudes toward actors, plot, and visual effects, and guides future work on implicit opinions and broader sentiment domains.
Alibaba's X‑DeepLearning (XDL) is an open‑source deep‑learning framework optimized for high‑dimensional sparse data, offering industrial‑grade distributed training, built‑in CTR/recommendation algorithms, structured compression, and online learning capabilities, with benchmark results demonstrating superior scalability and performance.
This article describes how a large‑scale operations monitoring system at Ctrip evolved from rule‑based alerts to a deep‑learning‑driven anomaly detection pipeline using LSTM models and Flink streaming, achieving a ten‑fold reduction in alert volume while improving fault detection.
This article explains the design of a Cross‑Domain Embedding (CSDE) method that transfers Alipay user features to Taobao representations, details its learning and adaptive prediction stages, and shows experimental and online results demonstrating significant conversion‑rate improvements for new and inactive users.
The article analyzes how iQIYI leverages deep learning, dynamic frame interpolation, personalized content features, and AI-driven heat metrics to reshape video streaming, improve traffic prediction, and gain a competitive edge in the rapidly evolving entertainment industry.
At Intel Architecture Day 2018 the company announced its next‑generation Sunny Cove micro‑architecture built on 10 nm, detailed its roadmap through 2021, introduced the 3D‑stacking Foveros technology, previewed new integrated and discrete GPUs, and launched a deep‑learning reference stack for AI developers.
The paper introduces an improved two‑channel CNN, built on a shared VGG16 backbone and a hinge‑loss metric, to automatically generate numeric tags for game advertising assets by learning content and style similarity, achieving over 97% test accuracy and enabling efficient ad placement and asset management.
AutoML and Neural Architecture Search automate deep‑learning model design by using controller networks to explore and evaluate candidate architectures, with efficient variants like PNAS and ENAS reducing cost, while platforms such as Google Cloud AutoML and open‑source AutoKeras make these techniques accessible, promising broader, democratized AI breakthroughs.
AutoML and Neural Architecture Search automate deep‑learning model design by sampling and training network blocks, using reinforcement‑learning or efficient weight‑sharing strategies such as PNAS and ENAS, enabling high‑accuracy architectures in days on a single GPU, with services like Google Cloud AutoML and open‑source tools like AutoKeras, while future research aims to expand search spaces beyond hand‑crafted blocks.
An in‑depth overview of multi‑task learning for natural language processing is presented, covering deep learning foundations, challenges, various multi‑task learning paradigms (hard, soft, shared‑private, function‑level, hierarchical, and search‑based sharing), benchmark platforms, and future research directions, illustrated with numerous diagrams.
The article examines YouTube’s large‑scale recommendation system, detailing its deep‑learning architecture, the challenges of scale, freshness and noise, and the design choices in candidate generation, ranking, data collection, and evaluation that together deliver over 70% of user watch time.
The article surveys the nineteen leading AI technologies—from natural language generation and speech recognition to digital twins and marketing automation—detailing their core functions, common use cases such as customer service, security, content creation, and the key vendors delivering each solution.
This article presents a multi‑source pointer network approach for generating concise, brand‑preserving product short titles in e‑commerce, detailing problem formalization, model architecture, extensive offline and online experiments, and demonstrating significant improvements over traditional truncation and seq2seq baselines.
The article summarizes Zheng Zeyu's presentation on deep learning fundamentals, the evolution and features of TensorFlow, and how AI techniques such as neural networks and Faster R-CNN are applied to address data challenges and enable intelligent fashion design and recommendation.
Alibaba's X-Deep Learning (XDL) framework, the first open‑source deep learning system designed for high‑dimensional sparse data, powers advertising, recommendation, and search workloads, delivering industrial‑scale performance and boosting revenue while offering an open, easy‑to‑use solution for the broader AI community.
This comprehensive guide walks beginners through fourteen practical steps to learn Python machine learning, covering essential Python skills, core scientific libraries, fundamental algorithms, advanced techniques like SVM and ensemble methods, dimensionality reduction, and deep learning with TensorFlow, all using free online resources.
This article explains how to speed up TensorFlow deep‑learning model training using single‑GPU acceleration, multi‑GPU parallelism, and distributed TensorFlow on Kubernetes, covering device placement, session parameters, synchronous vs asynchronous training modes, and practical code examples to improve performance and scalability.
This article discusses the limitations of algorithm-based classification models, including the need for large labeled datasets, limited sample coverage, frequent category changes requiring retraining, and complex optimization issues, while exploring knowledge graph-based approaches and generative adversarial networks for more flexible and accurate classification.
This comprehensive overview examines the fundamentals, historical milestones, key algorithms, major datasets, policy support, industry applications, and future trends of face recognition technology, highlighting its rapid growth within computer vision and artificial intelligence.
This guide lists and briefly describes 18 open‑source and cloud‑based machine learning platforms—from H2O and TensorFlow to Azure ML and AWS services—highlighting their key features, supported languages, and typical use cases for developers at any skill level.
This article presents a step‑by‑step solution that combines computer‑vision preprocessing, OCR, CNN classification, shape matching, and inpainting to isolate meaningful foreground elements from images with intricate backgrounds, discussing practical results, limitations, and code implementations.
This article explains how reinforcement learning can be modeled and deployed to enable intelligent, interactive product search on e‑commerce platforms, detailing problem definition, system architecture, training methodology, online results, and future research directions.
The article surveys ten leading open-source AI tools and frameworks—including TensorFlow, SystemML, Caffe, Apache Mahout, OpenNN, Torch, Neuroph, Deeplearning4j, Mycroft, and OpenCog—detailing their primary features, supported languages, hardware compatibility, and typical use cases for research and development.
The article explains how to replace keyword matching with deep‑learning document embeddings in Elasticsearch by applying PCA dimensionality reduction, indexing vectors using Lucene’s KD‑tree structures via a custom plugin, and leveraging FAISS‑style nearest‑neighbour techniques to achieve fast, semantically aware similarity search.
The article presents a comprehensive overview of Meituan's intelligent customer service system, detailing its evolution, the roles of QABot, TaskBot and ChatBot, the underlying NLP and knowledge‑graph technologies, model implementations such as DSSM and Seq2seq, and the impressive operational results achieved in food‑delivery and ride‑hailing services.
Researchers propose a Deep Feedforward Sequential Memory Network (DFSMN) for speech synthesis that matches BLSTM quality while using only a quarter of the model size and achieving four times faster inference, making it ideal for memory‑constrained, real‑time edge devices.
This article introduces Advanced LSTM (A‑LSTM), which linearly combines multiple past hidden states to overcome traditional LSTM's one‑step dependency, and demonstrates its application in utterance‑level speech emotion recognition, achieving a 5.5% accuracy improvement through attention‑based weighted‑pooling RNNs and auxiliary speaker and gender tasks.
During the World Cup, Youku used a frame‑rate‑up‑conversion algorithm that synthesizes intermediate frames by combining block‑based motion estimation, optical‑flow refinement, and a rule‑based plus deep‑learning fusion, enabling smooth 25→50 fps video and preserving fast‑moving objects such as a football.
The article clarifies that artificial intelligence encompasses machine learning, which in turn includes deep learning, and uses real‑world examples—from fraud detection and customer clustering to image recognition and language translation—to illustrate how these data‑driven models learn patterns, make predictions, and transform many industries.
This paper introduces a Deep Feedforward Sequential Memory Network (DFSMN) for statistical parametric speech synthesis that matches BLSTM quality with only a quarter of the model size and four times faster inference, making it ideal for memory‑constrained, real‑time IoT devices.
The article summarizes a QCon 2018 presentation by Luo Jingjia of Manbang Group, detailing how AI and deep‑learning techniques are applied to model, optimize, and recommend vehicle‑cargo matching in China’s massive highway logistics network, including data handling, neural‑network design, and practical challenges.
The article introduces convolutional neural networks, explains core concepts such as convolution, padding, stride, and pooling, demonstrates how to calculate output dimensions, and provides a step‑by‑step Keras example that builds, compiles, and trains a multi‑layer CNN for image classification.
This article describes how deep learning techniques, including convolutional neural networks and CAM visualizations, are applied to automatically evaluate and select high‑quality images for online travel agency ticket listings, improving user experience while reducing manual effort.
This tutorial explains how to extend a deep‑learning object detection project with OpenCV and Python to process live video streams, covering setup, command‑line arguments, model loading, frame‑by‑frame detection, drawing bounding boxes, FPS measurement, and performance tips for achieving real‑time results.
This article reviews the evolution of Alibaba’s e‑commerce search and recommendation technologies, detailing Taobao’s unique challenges, the shift from rule‑based retrieval to large‑scale machine learning, real‑time online learning, deep learning and intelligent decision‑making, and outlines future directions toward cognitive intelligence.
DeepInsight is a distributed, micro‑service‑based platform that provides end‑to‑end data exposure, multi‑dimensional visual analysis, and interactive debugging for TensorFlow models, turning opaque neural networks into transparent, controllable systems through real‑time visualizations, dynamic data sets, and integrated lifecycle management.
This article explores the challenges of OCR on warped document images, reviews traditional and deep‑learning‑based correction methods, describes a synthetic dataset generation pipeline, proposes enhanced U‑Net architectures including stacked and dilated variants, evaluates them with MS‑SSIM, and outlines future research directions.
JD.com's technical architecture team partnered with product teams to leverage the Euler knowledge computing platform and product knowledge graph to enhance the e-commerce review ecosystem, improving user experience in UGC communities through intelligent content management.
In an interview at the 2018 AI Pioneer Conference, Dr. Lv Zhengdong discusses his background, the development of neural‑symbolic AI at DeepCuriosity, its applications in law, public security and finance, challenges in NLP, and his perspective on the future trajectory of artificial intelligence.
The article surveys six emerging deep‑learning trends—capsule networks that retain spatial hierarchies, data‑efficient deep reinforcement and transfer learning, supervised models, memory‑augmented architectures such as long‑term and progressive networks, and hybrid Bayesian‑GAN approaches—highlighting how these advances expand AI capabilities beyond traditional fully‑connected networks.
The article details JD.com’s 3.5‑generation delivery robots, explaining their high‑precision multi‑sensor localization, deep‑learning perception, reinforcement‑learning control, extensive patent portfolio, and future challenges, while also inviting readers to vote for the robots in a national patent competition.
The article details JD's third‑generation autonomous delivery robots, covering their multi‑sensor fusion localization, deep‑learning perception, reinforcement‑learning motion control, extensive patent portfolio, and upcoming technical hurdles such as high‑precision mapping and lidar cost, while also inviting public voting for patent awards.
Online click‑through‑rate (CTR) prediction demands millisecond‑level response times, yet deep models are too slow; this paper introduces a “Rocket Launching” framework that jointly trains a lightweight net and a powerful booster net, sharing parameters and using gradient‑blocking and hint loss to improve accuracy without increasing inference latency.
The article explains that machine learning encompasses a range of algorithms such as decision trees and random forests, while deep learning—a specialized subset using multi‑layer neural networks—requires large data, powerful hardware, and longer training, yet offers superior performance in fields like computer vision, NLP, and medical diagnosis, and both are poised for expanding industrial and research adoption.
This article explains how Alipay combats the surge of mobile payment fraud by upgrading its risk‑identification system with deep‑learning techniques, modeling victim and fraudster behavior sequences using LSTM, and integrating the resulting scores into existing models to achieve a measurable increase in detection coverage.
This article introduces Wide&Deep, PNN, DeepFM, and a novel Wide‑ResNet model for Alibaba’s 1688 “Guess You Like” recommendation, explains the underlying feature services and real‑time scoring pipeline, presents offline experiments showing AUC gains with batch normalization, and shares practical tuning insights.
The talk surveyed game‑AI evolution—from rule‑based systems to AlphaGo‑style reinforcement learning—highlighted industry and academic methods, detailed challenges of applying deep‑learning techniques to MOBA titles like Honor Kings, and proposed a hierarchical, multimodal framework with analysis and execution modules supported by robust simulation environments.
This article surveys the evolution of face recognition from early statistical methods to modern deep‑learning approaches, outlines key researchers, open‑source projects, popular APIs, core processing steps, the DeepFace architecture, datasets, and experimental results, providing a comprehensive guide for practitioners and researchers.
The article explains how the black‑box nature of deep learning creates security risks such as adversarial attacks, describes real‑world examples in autonomous driving and medical imaging, and showcases JD Security's explainable AI system that demystifies model decisions to improve AI safety and industry adoption.
This article reviews deep‑learning‑based text classification techniques—including TextCNN, BiGRU, Capsule Networks, Attention mechanisms, and the novel cw2vec embedding—applied to Alipay complaint fraud data, presents experimental comparisons, and discusses their advantages, challenges, and future directions.
The article summarizes senior AI expert Wu Jianjun’s presentation on machine‑learning and deep‑learning engineering at Ping An Life, detailing the company’s big‑data platform, data processing pipelines, model training frameworks, distributed computing strategies, and production model‑serving architecture for financial applications.
The talk outlines OCR’s evolution from early postal-code readers to modern deep‑learning models, explains Tencent Cloud’s fast, accurate services for printed and handwritten text—including table‑structured and general OCR—and showcases real‑world applications such as ID cards, business cards, license plates, checks, and medical documents while highlighting ongoing challenges and future enhancements.
This article reviews a 2017 ACM MM paper that introduces a multi‑level similarity‑aware CNN (MSP‑CNN) for person re‑identification, detailing its siamese architecture, dual similarity constraints, multi‑task training, experimental results on CUHK03, Market‑1501 and CUHK01, and its advantages for large‑scale deployment.
The article describes Alibaba’s AI‑driven prediction platform for its smart‑customer‑service bots, detailing background, order and issue prediction capabilities, deployed products, underlying algorithms such as DeepFM, DCN, reinforcement learning, streaming computation, and the platform’s modular architecture that enables scalable, automated model management.
Tencent's AI Breast System combines mammography, pathology, MRI and ultrasound analysis using a multi‑scale, progressive TMuNet model that processes four views, learns from physician feedback, and delivers lesion localization, malignancy scoring and automated reports, achieving up to 92% sensitivity and reducing annotation time.
This paper introduces a style‑adversarial autoencoder that separates content and style latent variables, uses a multi‑class discriminator, and demonstrates superior image generation and data augmentation across MNIST, face, and text datasets tasks.
This paper introduces a spatio‑temporal autoencoder that uses 3D convolutions and a weight‑decaying prediction loss to automatically learn video representations for detecting abnormal events in real‑world traffic surveillance, outperforming previous methods on public benchmarks and a newly collected traffic dataset.
This paper introduces a novel multi‑level similarity‑aware CNN (MSP‑CNN) for person re‑identification, applying distinct similarity constraints to low‑ and high‑level feature maps, integrating classification and similarity losses in a multitask framework, and demonstrating superior performance on CUHK03, CUHK01 and Market‑1501 benchmarks.
Tencent’s intelligent machine‑learning platform achieved a world‑record by training AlexNet in 4 minutes and ResNet‑50 in 6.6 minutes on ImageNet, using large batch sizes, mixed‑precision, LARS optimization, hierarchical synchronization, gradient fusion, and pipeline I/O techniques to overcome accuracy and scalability challenges.
Alibaba’s DiDa platform leverages deep learning models—including CNN, LSTM, and DAN—to automatically generate fashion outfit pairings and descriptive copy, integrating a context graph for business rules, supporting real‑time personalization across e‑commerce, and demonstrating significant efficiency and CTR gains.
The JD.com Infrastructure team won the SIGIR 2018 e‑commerce competition by developing a multi‑level deep‑learning model that tackled severe class imbalance, vocabulary drift, and hierarchical labeling, boosting F1 from 0.74 to 0.84 and earning paper acceptance at the SIGIR conference.
This article describes Meituan's large‑scale AI‑driven image moderation platform, detailing deep‑learning based watermark detection, celebrity face recognition, pornographic image detection, and scene classification techniques, along with system architecture, data preparation, model evaluation, and deployment considerations.
This article presents Alibaba’s first brand‑level ranking system that personalizes product ordering by modeling user brand preferences with an enhanced Attention‑GRU, detailing feature engineering, model improvements, extensive offline experiments on a massive Tmall dataset, and a successful online A/B test that increased CTR, ATIP, and GMV.
This paper presents GXN, a generative cross‑modal feature learning framework that enhances image‑text retrieval by incorporating both high‑level semantic similarity and fine‑grained local matching through a three‑step Look‑Imagine‑Match process, achieving state‑of‑the‑art results on MSCOCO and Flickr30K.
The AVS Working Group, backed by iQIYI, has launched the Deep Learning Based Video Coding Loop Filter Challenge to spur research on AI‑driven loop filters that can cut bitrate by over 15%, inviting participants to submit Caffe models on the TAVS3 platform with ongoing benchmarking and potential prizes, while encouraging contributions to the upcoming AVS video coding standard.
Meituan employs AI-driven image moderation across millions of daily uploads, using an SSD‑ResNet detector for watermarks, a multi‑scale Faster R‑CNN ensemble for celebrity faces, a CNN‑based pornographic classifier with incremental learning, and transfer‑learned scene classification, while routing uncertain cases to human reviewers.
This article details the engineering of MobileNet V1 for image classification on mobile terminals, covering its depthwise separable convolution architecture, data collection and preprocessing, model training with transfer learning, TensorFlow Lite conversion, deployment on iOS/Android, and GPU acceleration techniques for faster inference.
The article reviews the evolution of artificial intelligence, presents a comprehensive AI reference framework based on roles, activities and functions, explains the intelligent information chain and IT value chain, and details current AI technology trends such as machine learning, deep learning, transfer learning, active learning and evolutionary learning, while also noting talent shortages and promoting an AI education course.
This article explains online action prediction, the challenges of early‑stage classification, and introduces a Scale Selection Network that dynamically chooses optimal temporal windows using dilated convolutions, regression and classification sub‑networks, achieving state‑of‑the‑art results on two benchmark datasets.
This article reviews the background of Alibaba's XiaoMi chatbot, explores how machine reading comprehension can be applied to e‑commerce scenarios such as rule interpretation and product inquiries, surveys key datasets and SQuAD‑based models, and discusses practical challenges and solutions for deploying these technologies in real‑world business environments.
The article introduces JD AI's JDAI-Face system, a deep‑learning based real‑time multi‑task facial attribute recognition platform that detects gender, age, ethnicity, expression and attractiveness, outlines its technical pipeline, showcases retail applications, and cites recent academic publications and expert contributors.
This paper addresses the challenge of ranking heterogeneous data in e‑commerce by proposing two algorithms—a multi‑armed bandit approach and a personalized Markov deep neural network—to select and order content streams, demonstrating superior performance over baseline models in A/B tests.
This tutorial explains how to implement high‑accuracy face recognition using OpenCV, Python, and deep learning by leveraging dlib's deep metric learning, creating a custom dataset, encoding facial embeddings, and performing real‑time identification on images and video streams.
The paper introduces a context‑contrast local feature and a gated multi‑scale fusion mechanism that together enhance pixel‑level scene segmentation, especially for inconspicuous objects, and validates the approach with state‑of‑the‑art results on Pascal Context, SUN‑RGBD, and COCO Stuff datasets.
This beginner-friendly guide walks through core machine‑learning concepts—from data organization and feature design to supervised and unsupervised algorithms such as perceptron, logistic regression, decision trees, LDA, and ensemble techniques—while explaining model evaluation, overfitting, and practical tuning strategies.
Meituan leverages deep‑learning models such as Word2Vec, DSSM, and LSTM‑based encoders within its ClickNet framework to compute text similarity and rank results, integrating rich business features like user location and merchant rating, thereby surpassing traditional TF‑IDF, BM25, and XGBoost approaches and boosting click‑through rates and revenue.
In June 2018, JD AI Research Institute's Computer Vision and Multimedia Lab won both the single‑person and multi‑person tracks of the CVPR 2018 Look Into Person (LIP) competition, achieving 90.9% and 72.2% accuracy respectively through enhanced multi‑scale CNN models, data augmentation, and focal loss techniques.
This article explains how Alipay’s xNN deep‑learning engine tackles the challenges of deploying AI on billions of mobile devices by using aggressive model compression, a lightweight SDK, and joint algorithm‑ and instruction‑level optimizations to achieve high accuracy, tiny package size, and real‑time performance.
The 2018 KDnuggets survey of over 2,000 data‑science professionals shows Python dominating with 66% usage, R dropping below 50%, TensorFlow leading deep‑learning frameworks, RapidMiner gaining traction, SQL remaining stable, Hadoop declining, and regional participation shifting toward Europe.
Alibaba’s DAMO Academy unveiled the DFSMN speech‑recognition model, open‑sourced on GitHub, which sets a new 96.04% accuracy record on LibriSpeech, trains three times faster than LSTM, and powers real‑world demos like AI cashiers and metro ticket machines.
Alibaba's DAMO Academy has open‑sourced the Deep‑Feedforward Sequential Memory Network (DFSMN), a next‑generation speech‑recognition model that achieves a world‑record 96.04% accuracy on LibriSpeech, trains three times faster than LSTM, halves model size, and dramatically speeds up real‑time decoding.
This tutorial walks through rewriting the HED edge‑detection network for TensorFlow 1.7, covering deprecated API fixes, migration from TF‑Slim to tf.layers, matrix initialization, batch normalization nuances, and a comprehensive review of convolution variants such as 1×1, depthwise, separable, and dilated convolutions, plus guidance on transposed convolutions and modern architectures like ResNet and Inception.
To address the rising cost of manual UI testing on diverse mobile devices, the article presents a deep‑learning‑based solution using PaddlePaddle that automatically detects UI style bugs such as misaligned controls, text overlap, and blank spaces through data‑driven model training, image preprocessing, and classification.
This article analyzes MXNet's lack of native video I/O, compares existing image iterators, introduces a Rec‑format based video iterator, and demonstrates through single‑GPU and multi‑GPU experiments that the new approach can accelerate training by up to eighteen times.
The team from Jilin University, Harbin Institute of Technology, and Beijing University of Posts and Telecommunications shares their winning strategy for the Tencent Advertising Algorithm Competition, detailing their feature engineering, model selection, and future work to handle large‑scale data challenges.
In this interview, Long Mingkang, Vice President of iFlytek's Cloud Computing Institute, shares his experience building large‑scale speech cloud services, discusses the technical hurdles of speech recognition and AI platform development, compares TensorFlow and MXNet, and offers insights on AutoML, industry trends, and how engineers can master AI.
Judea Pearl, the 2011 Turing Award laureate, argues that modern AI is stuck in curve‑fitting and that true intelligence requires machines to understand cause and effect, a perspective he expands on through a series of insightful interview questions and answers.
