Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Dr. Shang-Ling Jui - HUAWEI_HiAI_Enabling_AI_fo...

Dr. Shang-Ling Jui - HUAWEI_HiAI_Enabling_AI_for_Everyone

droidcon Berlin

July 17, 2018
Tweet

More Decks by droidcon Berlin

Other Decks in Programming

Transcript

  1. Security Level: HUAWEI HiAI Enabling AI for Everyone Dr. Shang-Ling

    Jui, Chief Scientist of Artificial Intelligence Consumer Business Group, Huawei
  2. Security Level: Intelligent phone vendor Developer ecosystem Limited AI apps

    Many great ideas but limited skills to leverage AI capability in smartphone Mobile AI application status quo
  3. Security Level: Low performance efficiency Insufficient computing power Limited data

    model Insufficient resources High Barrier Data collection Training optimization User experience optimization Long development period Low Efficiency Fragmented user experience Fragmented Capabilities Single scenario Incoherent user experience Hard to upgrade model Hard to obtain valid data Difficult Improvement Difficult capability enhancement Challenges for developers
  4. Service capability openness — Win-Win Pushes services to users based

    on users' needs, enabling services to proactively find users. App capability openness — Simple Integrates multiple AI capabilities to create more intelligent apps. Chip capability openness — Efficient Converts and migrates existing models to obtain an optimal performance using heterogeneous scheduling and NPU acceleration. Cloud HUAWEI HiAI Engine HUAWEI HiAI Service Device HUAWEI HiAI Foundation Chipset HUAWEI HiAI & Partners Create AI Ecosystem
  5. • Support 15 AI APIs • Highly efficient offline mode

    • Support multiple frameworks:Caffe, TensorFlow and TensorFlow Lite • Cover 90+ operators HUAWEI HiAI Foundation HUAWEI HiAI Engine CV Engine NLU- Engine ASR-Engine TensorFlow / TensorFlow lite/ Caffe etc HiAI Foundation DDK / Android NN HiAI Heterogeneous Resource Management / Runtime CPU GPU DSP NPU HUAWEI HiAI IDE HUAWEI HiAI Service HUAWEI HiAI Platform Leads the Way for AI Technology
  6. Real time Efficient Agile Provides acceleration APIs regardless of processor

    hardware to accelerate the model and operator calculation on mobile devices. Provides AI model management APIs such as model compilation, model loading, model running, and model unloading. Provides basic operator computing APIs, including interfaces such as convolution, pooling, and full connection. Benefit for developer: Fully Explore the NPU Computing Capability
  7. Night shot Dog Snow Flowers Greenery Cat Food Text Portrait

    Blue Sky Beach Sunrise /Sunset Stage performance Fireworks Group photo Document scan Waterfall Natural colors Close-up 500+ scenarios in 19 categories Benefit for customer: NPU Support Makes Devices More Intelligent
  8. * Data source: Huawei lab. Style suggestion based on HUAWEI

    HiAI Engine scene detection API 50X increase in performance effect ratio 20X increase in efficiency 2.5s single frame rendering Benefit for customer: NPU Support Makes Devices More Intelligent
  9. • Only two lines of code • JSON and Java

    are returned • 10–120 minutes public float getScore(Bitmap bitmap) throws RemoteException { if (bitmap == null) { Log.e(LOG_TAG,"bitmap is null "); return -1; } //Log.d(LOG_TAG,"runAestheticsService " + "define frame"); //demo by yang junyuan Frame frame = new Frame.Builder().setBitmap(bitmap).build(); Log.d(LOG_TAG,"runAestheticsService " + "start get score"); SparseArray<AestheticsScore> aestheticsScores = aestheticsScoreDetector.detect(frame); AestheticsScore score = aestheticsScores.get(0); return score.getScore(); } Benefit for developer: Easy-to-use AI Capabilities Help to Accelerate App Development
  10. Live broadcasting Social platform Photo taking E-commerce Text processing Short

    Videos and Live Broadcasting Social Platform AR Photo Taking and Image Retouching Shopping Translation and Text Processing • Human face recognition • Gestures recognition • Portrait segmentation • Human posture identification • Video style configuration • Voice control • Intelligent DoF control • Image scenario identification • Photo classification • Image recognition • Image super- resolution • Sensitive information identification • Scenario identification • Voice control • Depth estimation • Light estimation • Beautification • Image enhancement • Aesthetics scoring • Album generation • Voice controlled photo taking • Gesture controlled photo taking • QR code scanning • Instant access and recommendation • ID card identification • Bank card identification • Shopping by scanning a picture • Translate by taking a photo • OCR • Word segmentation • Naming entity identification • Text emotion recognition • Intelligent reply • Text and image super-resolution CV and ASR CV and NLU ASR and CV CV CV NLU, CV, and ASR Benefit for customers: Easy-to-use AI Capabilities Help to Enrich App Scenarios
  11. Photo highlight with aesthetics score Benefit for customer: Easy-to-use AI

    Capabilities Help to Create More Natural Interaction
  12. 12

  13. Registered developers Users Huawei phone shipments Revenue sharing 350,000 340

    million 153 million(2017) 45 million USD Benefit for developers: An All-win Service Platform
  14. Cover 50+ categories of services 80% common scenarios 3 steps

    in average 5s to 10s for each complete service Most scenarios covered: Reduced interactions: Saved time: Best coverage: Benefit for customers: Easy-to-use Services
  15. • Log developers' habits. Easy-to-use development platform • Integrates AI

    models, usage wizard, and all-day remote one-click real device debugging • Quick, convenient, and efficient help available for developers to use Huawei HiAI Foundation and Engine capabilities. • Quickly develop brand new AI apps that expand the imagination. Comprehensive Tools Support, Perfect Integration of HUAWEI HiAI IDE and Android Studio
  16. 1. Select an appropriate API at service layer. 2. Reference

    to the API. Developers without self-owned AI models 1. Use the AI API at the app layer based on service requirements. 2. No self-owned AI model training and deployment involved Development process A1. Model development A2. Model training A3. Model cropping output B1. Model format conversion B2. Model import project B3. Model referenced code generating B4. Business logic invoking B5. SDK packing and deployment to the device A. Model training  Role: AI expert and data expert  Tools and framework: Tensorflow, Caffe, Torch, CNTK, Keras, scikit- learn, and more B. Model utilization  Role: App development engineer  Tools and framework: IDE, model conversion tool, and platform SDK Development process Developers with self-owned AI models 1. Self-owned AI training capability 2. Self-owned AI model, developed at the framework layer • Hour-level integration of HiAI Engine capability • Day-level model integration of HiAI Foundation Simple and Easy to Use
  17. Accelerated AI capability Mature service platform HiAI Foundation: NPU capability

    HiAI Engine: Easy-to-use AI capability HiAI Service: Service platform HiAI IDE: Perfect integration of Android Studio Comprehensive tools support AI capability provider Achieve win-win with developers
  18. HUAWEI HiAI Continuously Assists Developers to Achieve A Win-win Outcome

    Recommendations through Huawei official channels Home page recommendation in AppGallery AI category AppGallery Promotion in Huawei's 10,000 stores HUAWEI Store recommendation Special column recommendation through various Huawei official social network accounts Stickie on Huawei fans club Joint marketing Joint PR planning and promotion
  19. Model Size 50-70 MB 2 MB Processing Speed DDK* 5.1

    sec 1.7 sec Success Story – Prisma
  20. “One of the Best Mobile Office Suites” “A seriously impressive

    and attractive suite of tools.” “The most attractive Office 365 alternative…” Total Users Worldwide MAU 250 Millions Installs 1.25 Billions
  21. • Turn your images into a Presentation with a single

    click. • Easily preview, enhance, crop, and rotate to fine-tune your images. Convert to PPT One-Click Generation of Presentations
  22. 499 622 Duration of use in seconds Opened Documents Start-ups

    +123s 3.6 6 4.0 1 +0.35 5.7 4 10. 1 +4.36