2018 Suzhou GTC Conference Notes

Keynote

NVIDIA founder and CEO Jensen Huang’s keynote on the future of computing. Speaker: Jensen Huang, NVIDIA founder and CEO. 2018/11/21 Wed 10:00 - 12:00 | Main Hall, 3rd Floor, Jinji Lake Hall

Interpretability, Robustness, and Fairness: THUIR Personalized Recommendation Research [CH81402]

Recommendation systems play an ever larger role in daily life and work—e-commerce, social media, and news feeds all depend on personalization. Huge user bases challenge systems with massive interaction data; personalization also brings sparsity and missing data. This talk covers recent THUIR work on deep learning for recommendations: interpretability, user satisfaction, and unified models blending collaborative filtering and content for cold start. Published at WWW 2018, SIGIR 2018, CIKM 2018, and similar venues.

Speaker: Min Zhang, Tsinghua University, tenured associate professor

2018/11/21 Wed 13:30-14:00 | Breakout 7 (Room A204-205)

Interpretability, robustness, fairness—why these three: the first two are common AI issues; fairness grows as bias increases. Interpretability—for users, why this result; for engineers, why this recommendation works (algorithm, user matrix, item matrix, why this item scores high). Add latent variables like screen size to give users a reason; joint learning to improve clicks.

Surface useful reviews. 2018 paper. Matthew effect on reviews—later good reviews invisible = unfairness; surface new good reviews. Sometimes interpretability hurts fairness. Can interpretability be generated?

Robustness: imbalanced samples, few samples.

Cold start: collaborative filtering fails. Content-based and collaborative filtering become user-based and item-based.

Fairness: systems are user–item interaction; technical unfairness always exists; some items never get recommended. Diversity: increases item fairness, lowers user fairness. Quality: low-quality news gets better click-through. Method: focus on users, not click-through alone.

A personal philosophy problem and an algorithm problem.

Machine Learning Development and Industry Applications [CH8502]

Speaker: Yong Qin, IBM China, data scientist

2018/11/21 Wed 14:00-15:00 | Breakout 5 (Room A108-109-110)

IBM insurance similarity analysis; children learning words—recommend next item; “thinking” formal wear; autonomous vehicles; energy and environment management.

Medicine: interpretability, personalized intervention. Patient clustering with decision trees, mining, feature similarity. Medical data: missing data—KNN for similar cases, impute from similar records.

ML cannot easily cross domains.

Accelerating AI with TensorFlow [CH8203]

Recent AI and ML breakthroughs have spawned new applications; many research areas are still nascent. ML platforms are essential to speed new algorithms and production deployment. TensorFlow is open source, easy to use, flexible from research to production; its architecture deploys across desktops, clusters, mobile, and edge. Originally from Google Brain under Google AI for ML and deep learning; its numerical core serves broad science. Mike Liang covers TensorFlow applications and latest features for accelerating AI.

Speaker: Mike Liang, Google AI senior product manager

2018/11/21 Wed 15:00-16:00 | Breakout 3 (Room A104-105)

TensorFlow data input, batch input, online input, rd.graph, model serving. Mobile: converter.

Accelerate browser, Node.js.

GPU-Powered AI for Entertainment [CH81406]

PART I: GPU as driver of the AI wave. PART II: GPU empowering Yi+ AI: (1) matrix ops speed model tuning; (2) larger VRAM enables deeper models—top international vision competition results. PART III: Yi+ AI for entertainment: Camera + AI; watch-and-buy content commerce; watch-and-X (AI assistant, content moderation); contextual marketing (trends, TV ad bottlenecks, scene/data/algorithm/experience, advantages). Yi+ builds open API/SDK on three core engines for intelligent video/image and contextual ad platforms.

Speaker: Mo Zhang, Beijing Moshanghua Technology, founder and CEO

2018/11/21 Wed 15:40-16:10 | Breakout 7 (Room A204-205)

AI + entertainment.

AI + camera: category recognition, outfit change, more engaging interaction, dynamic AR camera.

Watch and buy: live TV, recognize and purchase. Large-screen AI assistant.

TV + AI: who is this star, content recommendation, what animal is this, auto editing, outfit change. Content moderation platform.

OTT content marketing: VIP cuts ad revenue; new ad ideas, more placements. Variety show ads without post-production—use AI.

Efficient Neural Machine Translation [CH8204]

GPU-driven neural MT beats statistical MT and is mainstream, but challenges remain: (1) needs large parallel data, low sample efficiency; (2) online latency lags statistical MT. Microsoft Research Asia shares: (1) monolingual unlabeled data for accuracy and efficiency; (2) non-autoregressive decoding for speed.

Speaker: Tao Qin, Microsoft Research Asia, senior researcher / manager

2018/11/21 Wed 16:00-17:00 | Breakout 3 (Room A104-105)

Machine translation: dual learning for missing labels—like reinforcement learning, lower annotation cost. Unsupervised: two models. Transfer learning: model from one task to another; dual supervised learning with regularization.

Scaling AI Research to Production [CH8205]

Facebook’s AI edge comes from strong infrastructure bringing cutting-edge research to large-scale production. PyTorch 1.0 bridges imperative and declarative modes to speed research-to-production. FBLearner software and Big Basin hardware support large-scale ML.

Speakers: Xiaodong Wang, Facebook research scientist; Yiming Wu, Facebook software engineer

2018/11/21 Wed 17:00-18:00 | Breakout 3 (Room A104-105)

Facebook: applying research to production.

Recommendations, machine translation, image-to-speech, assistants, auto titles for user videos, AR. Phishing post detection, suicide post detection, help. Ad delivery and ranking. Fake ad user removal.

Steps: trained models need conversion for deployment efficiency. PyTorch for research and freedom; Caffe2 for products; ONNX for conversion. PyTorch 1.0 upgrade. Fast academia–industry transfer. What is industrial-grade ML framework? Hardware, scale, cross-platform.

New features: (1) conda integration; (2) tensor optimization; (3) graph-level optimization like TensorFlow; (4) model performance, RT conversion; (5) distributed training; (6) network formats, c10d; (7) C++ deployment—Python model to C++; (8) quantization, low-precision; (9) mobile neural deployment.

Coding: static Caffe, graph mode. Dynamic Torch, direct edits.

1.0 combines both—eager code.

Intelligent Computing: Breaking Data Center Boundaries [CH8107]

Huawei intelligent computing breaks chip, server, and data center boundaries to deliver tiered compute for AI and critical business breakthroughs.

Speaker: Yu Liu, Huawei IT hardware chief architect

2018/11/22 Thu 10:00-11:00 | Breakout 8 (Room A206-207-208)

Huawei AI challenges: data collection, labeling, algorithm design, training and evaluation.

Vision Systems, Algorithms, and Optimization for Unmanned Retail [CH8608]

JD’s unmanned vending cabinet practice: NVIDIA Jetson TX2 algorithm selection and optimization, frontier exploration.

Speakers: Zhuoran Xu, JD algorithm researcher; Lishan Xiao, JD unmanned cabinet lead

2018/11/22 Thu 11:00-12:00 | Breakout 9 (Room A216-217-218)

JD unmanned retail partnership. Algorithms, optimization. Approaches: RFID tags and scanners for item type and count; weight sensors—slot for type, weight for count; static vision—photos, cloud processing, high accuracy, strict placement; dynamic vision—external cameras, heavy compute; dynamic vision + weight.

When AI Meets Games [CH8805]

Survey of AI’s impact on game dev in two years; academic AI vs game AI. ViZDoom 17–18: reinforcement learning for game AI, challenges, GPU experience in DRL training.

Speakers: Cheng Ge, Tencent expert engineer; Jia Xu, Tencent AI Lab expert researcher

2018/11/22 Thu 14:00-15:00 | Breakout 7 (Room A204-205)

Tencent games: action and environment adaptation; quadruped locomotion complexity. RL + physics, skill animation.

Facial animation—direct generation. Art asset generation: five-layer CNN, structured texture generation, less manual editing.

Game AI: human-like NPC behavior.

GPU-Accelerated Time-Series Query and Analysis [CH8111]

[Image upload failed…(image-b56b2b-1542953976139)]

Time-series databases are widely used in finance, internet, IoT, etc. CPU architectures struggle with massive time-series analytics. GPU parallel approach greatly improves performance: (1) time-series background, query process, challenges; (2) GPU design and key techniques; (3) performance comparison; (4) follow-up work.

Speaker: Xinyang Yu, Alibaba senior expert

2018/11/22 Thu 15:00-16:00 | Breakout 8 (Room A206-207-208)

Time-series DB: records, collection with defined intervals, use cases—analysis, diagnosis, prediction.

Example: network zero causes downstream issues. Weibo wedding-day crash—predict beforehand, alert, scale up.

Data scale: huge retention—TB per day; real-time courier locations. Framework: compression, inverted index, old data on disk; decompress 0.2 time, aggregate 0.7 time. How GPU decompresses and aggregates—group consecutive points.