infoShare AI Roadshow 2018 - Tomasz Kopacz (Microsoft) - jakie możliwości daje platforma Microsoft do budowy inteligentnych rozwiązań. I czy na pewno humaniści wyginą?
•
1 gefällt mir•80 views
Podczas tej sesji przyjrzymy się, w jaki sposób można skorzystać z platformy Microsoft do budowy tzw. „inteligentnych” rozwiązań. W przykładach zobaczymy zarówno Cognitive Services, jak i wykorzystaniu GPU (a dokładniej – Batch AI) do uczenia sieci neuronowych. Zajmiemy się także skomplikowanym zagadnieniami związanymi z projektowaniem – tak by algorytmy rozszerzały ludzkie możliwości (a nie nas zastępowały). Sesja zakłada że słuchacze umieją programować.
Empfohlen
Empfohlen
Weitere ähnliche Inhalte
Was ist angesagt?
Was ist angesagt? (12)
Ähnlich wie infoShare AI Roadshow 2018 - Tomasz Kopacz (Microsoft) - jakie możliwości daje platforma Microsoft do budowy inteligentnych rozwiązań. I czy na pewno humaniści wyginą?
Ähnlich wie infoShare AI Roadshow 2018 - Tomasz Kopacz (Microsoft) - jakie możliwości daje platforma Microsoft do budowy inteligentnych rozwiązań. I czy na pewno humaniści wyginą? (20)
Mehr von Infoshare
Mehr von Infoshare (20)
Kürzlich hochgeladen
Kürzlich hochgeladen (20)
infoShare AI Roadshow 2018 - Tomasz Kopacz (Microsoft) - jakie możliwości daje platforma Microsoft do budowy inteligentnych rozwiązań. I czy na pewno humaniści wyginą?
- 3. What is AI, ML, DNN, …. and how we (will) build nextgen software Cognitive Services Why Python ☺ Azure Machine Learning Services
- 5. 1950 1960 1970 1980 1990 2000 2010
- 6. Learn it when build formal, classic algorithm is too complicated
- 8. Future of Software Program Logic + Artificial Intelligence
- 9. SW Development AI Development Data Schemas Data Curation + Data Labeling Functions & Code Machine Learning Models Debugger to process A/B Testing, Experimentation Use telemetry to inform Telemetry driven learning
- 13. Input Output
- 15. 3 Error back propagation Feedforward of information
- 16. Machine Learning & AI Portfolio When to use what? What engine(s) do you want to use? Deployment target Which experience do you want? Build your own or consume pre- trained models? Microsoft ML & AI products Build your own Azure Machine Learning Code first (On-prem) ML Server On- prem Hadoop SQL Server (cloud) AML services SQL Server Spark Hadoop Azure Batch AI DSVM Kubernetes Visual tooling (cloud) AML Studio Consume Cognitive services, bots
- 18. Computer Vision Face/Emotion Recognition OCR/Handwriting Custom Vision Video Indexer Content Moderator Text-to-Speech Speech-to-Text Translator Custom Speech Language Understanding PII Detection Text Translator Text Analytics QnA Maker Bing Custom Search Bing Visual Search
- 25. IEEE, 2017
- 27. Sophisticated pretrained models To simplify solution development Azure Databricks Machine Learning VMs Popular frameworks To build advanced deep learning solutions TensorFlow KerasPytorch Onnx Azure Machine Learning LanguageSpeech … Azure Search Vision On-premises Cloud Edge Productive services To empower data science and development teams Powerful infrastructure To accelerate deep learning Flexible deployment To deploy and manage models on intelligent cloud and edge Cognitive Services
- 29. Top 20 Python libraries for data science in 2018 NumPy SciPy Pandas StatsModels Matplotlib Seaborn Plotly Bokeh Pydot Scikit-learn XGBoost LightGBM CatBoost Eli5 TensorFlow PyTorch Keras Dist-keras elephas spark-deep-learning NLTK SpaCy Gensim Scrapy Microsoft Cognitive Toolkit (CNTK) Caffe2
- 30. Developer Data Scientist (Feature Designer / domain expert) And: someone working with „human species”
- 34. My Computer Data Store Azure ML Workspace Compute Target Docker Image
- 35. Prepare Data Register and Manage Model Train & Test Model Build Image … Build model (your favorite IDE) Deploy Service Monitor Model Prepare Experiment Deploy
- 37. Virtual Machine • New in Azure | Existing in Azure • Anywhere (on prem as well!) Azure Batch AI • Best choice regarding price/performance • + Low priority instances Azure Kubernetes Service (AKS) • Maybe for CPU bound? Many parallel calculations? Existing „free” cluster?
- 39. Pascal generation GPU instances - NVIDIA Tesla P100 GPUs Targeted for accelerating machine training jobs and HPC Specs: FP64 – 4.7 TFLOPS of double precision floating point performance FP32 – 9.3 TFLOPS of single precision performance FP16 – 18.7 TFLOPS of half-precision performance GPU Memory 16 GB NC6s_v2 NC12s_v2 NC24s_v2 NC24rs_v2 Cores (Broadwell 2.6Ghz) 6 12 24 24 GPU 1 x P100 2 x P100 4 x P100 4 x P100 Memory 112 GB 224 GB 448 GB 448 GB Local Disk ~700 GB SSD ~1.4 TB SSD ~3 TB SSD ~3 TB SSD Network Azure Network Azure Network Azure Network Azure Network + InfiniBand
- 40. Volta PCIe GPU instances - NVIDIA Tesla V100 GPUs Targeted for accelerated machine training jobs and HPC Specs: FP64 - 7 TFLOPS FP32 - 14 TFLOPS GPU Memory 16 GB NC6s_v2 NC12s_v2 NC24s_v2 NC24rs_v2 Cores (Broadwell 2.6Ghz) 6 12 24 24 GPU 1 x V100 2 x V100 4 x V100 4 x V100 Memory 112 GB 224 GB 448 GB 448 GB Local Disk ~700 GB SSD ~1.4 TB SSD ~3 TB SSD ~3 TB SSD Network Azure Network Azure Network Azure Network Azure Network + InfiniBand
- 41. Pascal generation GPU instances - NVIDIA Tesla P40 GPUs Targeted for machine training and inference jobs Specs: FP32 – 12 TFLOPS INT8 - 47 TFLOPS GPU Memory 24 GB ND6s ND12s ND24s ND24rs Cores (Broadwell 2.6Ghz) 6 12 24 24 GPU 1 x P40 2 x P40 4 x P40 4 x P40 Memory 112 GB 224 GB 448 GB 448 GB Local Disk ~700 GB SSD ~1.4 TB SSD ~3 TB SSD ~3 TB SSD Network Azure Network Azure Network Azure Network Azure Network + InfiniBand
- 42. Volta SXM GPU instances - NVIDIA V100 GPUs 8X NVIDIA V100 GPUs interconnected with NVLink mesh Excellent for accelerating machine training jobs and HPC Skylake based processor with premium storage support (SSD backed) Availability: Q4 CY2018 Specs: GPU Memory 16 GB 300 GB/s GPU interconnect through NVLink ND40s_v3 Cores 40 cores GPU 8 x V100 SXM Memory 672 GB Local Disk ~1.3 TB SSD Network Azure Network + NVLink GPU interconnect
- 43. NV6 NV12 NV24 Cores (Haswell 2.6 GHz) 6 12 24 GPU 1 x M60 (half card) 2 x M60 (full card) 4 x M60 Memory 56 GB 112 GB 224 GB Local Disk ~340 GB SSD ~680 GB SSD ~1.4 TB SSD Network Azure Network Azure Network Azure Network GRID Licenses 1 2 4 NV6s_v2 NV12s_v2 NV24s_v2 Cores (Broadwell 2.6Ghz) 6 12 24 GPU 1 x M60 (half card) 2 x M60 (full card) 4 x M60 Memory 112 GB 224 GB 448 GB Local Disk ~700 GB SSD ~1.4 TB SSD ~3 TB SSD Network Azure Network Azure Network Azure Network GRID Licenses 1 2 4 GPU instances featuring NVIDIA Tesla M60 GPUs Haswell & Broadwell based CPU processor Premium storage support on NV_v2 (SSD backed) Grid license included with each GPU instance 1 workstation/GPU or 25 vApps/GPU Specs: ~8TF FP32 per full card 36 H.264 1080p30 streams GPU Memory 8GB GDDR5 GB – 16 GB per cardhttps://aka.ms/nvv2signup
- 44. Applications GPU Provisioning Host OS Client OS Hardware • Azure Developer & Platform Services • Custom Images • Azure Marketplace • Custom apps and services • Driver • Hyper-V • Discrete Device Assignment (DDA) • NVIDIA GPU
- 47. Neural Functional Unit VRF Instruction Decoder TA TA TA TA TA Matrix-Vector Unit Convert to msft-fp Convert to float16 Multifunction Unit xbar x A + VRF VRF Multifunction Unit xbar x + VRF VRF Tensor Manager Matrix Memory Manager Vector Memory Manager DRAM x A + Activation Multiply Add/Sub Legend Memory Tensor data Instructions Commands TA Tensor Arbiter Input Message Processor Control Processor Output Message Processor A Kernel Matrix Vector Multiply VRFMatrix RF + Kernel Matrix Vector Multiply VRFMatrix RF Kernel Matrix Vector Multiply VRFMatrix RF NetworkIFC ...
- 50. Microsoft Cognitive Toolkit TensorFlow Chainer
- 71. Python Script High Quality Machine Learning Model Dataset Automated ML Models (User Compute – Local or Cloud) Output Generate Algorithms & Hyperparameter values
- 74. Number_of_ hidden_layers Number_of_ nodes_in_layer Input layer Hidden layer 1 Hidden layer 2 Output layer E.g. Learning_rate Batch_size
- 75. Challenges • Huge search space to explore • Sparsity of good configurations • Expensive evaluation • Limited time and resources Example Problem • Find “best” values for Number_of_layers and learning_rate • Number_of_layers – [2, 4, 8] • learning_rate – anywhere between 0 and 1 • Optimize for model accuracy
- 76. Dataset Training Algorithm Hyperparameter Values – config 1 Model 1 Hyperparameter Values – config 2 Model 2 Hyperparameter Values – config 3 Model 3 Model Training Infrastructure
- 78. #1 #2 #𝑝 𝑟𝑢𝑛1 𝑟𝑢𝑛2 𝑟𝑢𝑛 𝑝 … Hyperparameter Tuning runs in Azure ML 𝑟𝑢𝑛𝑗 𝑟𝑢𝑛𝑖 𝑟𝑢𝑛 𝑘 (B) Manage resource usage of active runs • How long to execute a run? ??? (A) Generate new runs • Which parameter configuration to explore? 𝒓𝒖𝒏 𝟏= {learning rate=0.2, #layers=2, …} 𝒓𝒖𝒏 𝟐= {learning rate=0.5, #layers=4, …} 𝒓𝒖𝒏 𝒑= {learning rate=0.9, #layers=8, …}
- 79. { “learning_rate”: uniform(0, 1), “num_layers”: choice(2, 4, 8) … } Config1= {“learning_rate”: 0.2, “num_layers”: 2, …} Config2= {“learning_rate”: 0.5, “num_layers”: 4, …} Config3= {“learning_rate”: 0.9, “num_layers”: 8, …} … Sampling algorithm
- 80. Early terminate poorly performing runs
- 84. “We believe AI technology has the power to amplify human ingenuity and extend our capabilities so we can achieve more. When made accessible to everyone, AI will transform industries, make us more productive, and help solve society’s biggest challenges. This intelligent technology is already improving our lives today and will change the world tomorrow in ways unimaginable to us now.
- 85. - Satya Nadella “The most critical next step in our pursuit of A.I. is to agree on an ethical and empathetic framework for its design.”
- 86. 3 principles of inclusive design 1 2 3 Recognize Exclusion exclusion happens when we solve problems using our own biases Learn from Diversity human beings are the real experts in adapting to diversity Solve for one, Extend to many by focusing on what's universally important to all humans
- 89. Designing for human diversity
- 91. A young child defines the world purely on the small amount they can see. Eventually, the child learns that most of the world lies beyond the small set of information that’s within their field of vision. Applications of AI Face recognition Speech recognition Image recognition Risk of exclusion Gender unbalanced Narrow age range Race and ethnic groups • Does that sample include everyone in your customer base? • Have you tested your results with people who weren’t part of your sample? • Is your AI team inclusive, diverse, and sensitive to recognizing bias? Stress test Dataset bias Data doesn’t represent the diversity of the customer base
- 92. Dataset bias Nikon Camera blink recognition Hewlett-Packard web cam face framing Training App for Face recognition
- 93. Imagine some kids who like to play “doctor.” The boys want the doctor roles and assume the girls will play the nurses. The girls have to make their case to overturn assumptions. “Hey, girls can be doctors too!” Applications of AI Search engines Data sources and history Translation and autocorrect Risk of exclusion Gender discrimination Race / ethnic discrimination Political / religious views • Are your results making associations that perpetuate stereotypes? • What can you do to break undesirable and unfair associations? • Is your dataset already classified and labeled? Stress test Associations bias Data associations reinforces and multiplies cultural bias
- 94. Associations bias Google Re captcha testGender unbalance translations Google Image search results
- 96. But – another approach - Robot Vera
- 97. Imagine a girl getting a makeover. The girl likes sports, loves a natural look and hates anything artificial. The beautician has different ideas about beauty, applies tons of makeup and a fussy hairdo. Applications of AI Algorithmic decision making Big data and privacy Self-driving cars Risk of exclusion Justice and Fairness Financial Services Hiring and Human Resources • Would real, diverse customers agree with your algorithm’s conclusions? • Is your AI system overruling human decisions and favoring automated decision? • How do you ensure there’s a human POV in the loop? Stress test Automation bias When automated decisions override social and cultural considerations.
- 98. Automation bias Face App, attractiveness filter https://github.com/Microsoft/AirSimAmazon one-day delivery
- 99. A popular kids’ game is “Telephone.” The first person in a group whispers a sentence to next person, who then whispers it to the next person . But say one kid changes it intentionally to create a more ridiculous result. Applications of AI Business and social chat bots Digital personal assistants Service / entertainment robots Risk of exclusion Digital social competences Emotion recognition Personality and embodiment • Did you design for real-time interaction and learning? • Do you have checks in place to identify malicious intent toward your system? • What does that mean for what it reflects back to customers? Stress test Interaction bias When humans tamper with AI and create biased results.
- 100. Interaction bias Tay social chat bot Google Perspective API for toxicityFrequently bought together
- 101. Applications of AI Social networks / messaging Recommendations / ratings News and comments Risk of exclusion Stereotyping Echo Chamber Fake News • Does your algorithm build on and reinforce only popular preferences? • Is your AI able to evolve dynamically as your customers changes over time? • Is your AI helping your customers to have a diverse and inclusive view? Stress test Confirmation bias When oversimplified personalization makes biased assumptions and narrows customers views. Think of the kid who gets a toy dinosaur for a present one year. In several years, friends and family assume the kid is a dinosaur fanatic, and keep giving more dinosaurs until he has a huge collection.
- 102. Confirmation bias Recommendations after search Recommendations after purchase Tracking prevention
- 103. 3 principles of inclusive design 1 2 3 Recognize Exclusion exclusion happens when we solve problems using our own biases Learn from Diversity human beings are the real experts in adapting to diversity Solve for one, Extend to many by focusing on what's universally important to all humans