使用AI服务作用于一种交付设备

惠普投资了一种新的产品交付模式,称为设备即服务(DAAS)。DaaS投资的成功取决于设备的交付、监视、替换、用户交互和服务的自动化。DaaS的核心是一组虚拟助理,通过积极的成本模型优化成本和使用经验,确保客户满意。这次演讲的关键是惠普如何利用开发虚拟助理来改变工作场所。此外,约翰还将介绍惠普在SPARK上开发AI的方法,以及惠普为什么选择SPARK作为AI的核心技术。
展开查看详情

1.USING AI TO DELIVER A DEVICE AS A SERVICE Franco Vieira, HP Inc. #Ent1SAIS

2.Agenda • HP and AI • Device Telemetry • Case Study • Next Steps

3. TWO Personal Printing leading Systems franchises 3

4. DATA 4

5.Device Telemetry What Comes With This Data? Are your customers using the features you expect? CONTEXT How are they engaging with your product? How frequently are users engaging with your service, and for what duration? What settings options to users select most? Do they prefer certain display types, input modalities, screen INSIGHTS orientation, or other device configurations? What happens when crashes occur? Are crashes happening more frequently when certain features are used? What’s the context surrounding a crash?

6.Device Telemetry Application Areas Improving Existing Services for Products and Individuals and New Experiences Experiences Fleets

7.Device Telemetry Challenges • Gather information for hundreds, perhaps thousands of devices • Megabytes, perhaps gigabytes of data per hour coming off a single device • Massive amounts of data will drive data governance Source: David Linthicum, The Data Challenges of Telemetry, 2014

8.Case Study Health Management

9.A Unified Approach to Analytics

10. Machine Learning Pipeline Training Pipeline Selecting Data Transformation Dataset Preparation Training Telemetry • Filtering • Windowing • Feature Extraction • Normalization • Labeling • Training models Data • Rescaling • Preparing data • LSTM • Outliers removal structures for training • CNN+LSTM • … Millions of Thousands of Thousands of Devices Devices Timeseries

11. Machine Learning Pipeline Inferencing Pipeline dist-keras Selecting Data Transformation Dataset Preparation Inference Telemetry • Normalizing • Preparing data • Loading models • Rescaling structures for inferencing • Running inferences Data • Windowing Millions of Millions of Millions of Devices Devices Predictions

12. Typical Results Unsupervised Anomaly Detection Unsupervised Grade Analysis Multiclass Classification

13.Closing The Loop

14. Next Steps • Technical – Advanced Automated Learning Pipeline – Improve Repeatability – Massive Parallel Processing • Business – Increase the number of AI services – Improve user experience

15.Franco Vieira HP Personal Systems Data Science franco.vieira@hp.com