Sponsors tinyML Committee tinyML Org team: Bette Cooper, tinyML - PowerPoint PPT Presentation

tinyML- 2019 Sponsors

tinyML Committee

tinyML Org team: Bette Cooper, tinyML Org Lead • Gary Brown • Gina Edwards and Ira Feldman Org. aspects and local Google team: arrangements • Edd Wilder-James • Marcus Chang • Daniel 'Wolff' Dobson “Let’s make tinyML BIG !”

Why tinyML ? Data generated: Data is a new oil(electricity) … and ML is a way to produce it Source: IDC Cloud ML 1% Storage and sharing • DNN on the cloud • HW: TPU, FPGA, GPU, CPU User provided: 4% Edge ML 1. Pics 2. Audio • Optimized algos and CNN-light 3. Clicks/likes • SoC (with NPUs/NSP accelerators) 4. GPS, Location based Tiny ML 95% • CNN-micro Real-time in the • MCU w/ HW accelerators physical world CMOS Audio Optrical IR IMUs Environ/ Temperature cameras micsb sensors cameras chemical

tinyML is “good enough” NOW … with more enhancements coming in the near future $$$ More tinyML apps and value creation $ initial tinyML applications SW HW Algos Quantization, compression HW accelerators (digital) Smaller models (100s kB) - Compute in memory - Novel algos/networks - Analog compute - 10s kB models - Neuromorphic Enabling technologies: ULP sensors, novel memories, 3D, energy scavenging, ULP radio

In the next 5 years TinyML can unleash over $70BN in economic value LOGISTICS AVERAGE $28BN CAGR MANUFACTURING / 2 7 . 3 % INDUSTRIAL AUTOMATION $22BN SMART CITIES/ BUILDING $12BN RETAIL $8BN The Silent IntelligenceResearch

tinyML Summit-2019 Objectives • Review the state of the art of tinyML • Identify gaps and opportunities, both tech and biz/products • Develop tinyML ecosystem/community and define future events Note: - tinyML-2019 focuses on technology aspects - future events will cover more applications, end-users, VC/funding, etc.

tinyML Summit Program format – March 20, 2019 • Two morning sessions of invited presentations on: • HW, moderated by Ian Bratt, ARM • Systems, moderated by Boris Murmann, Stanford Univ. • Long Lunch break and posters/demos/networking: 12:40-2:30 pm • Afternoon session on SW & applications, Kurt Keutzer, UC-Berkeley • Afternoon poster/demo session and networking: 4:30-6:00 pm • Dinner starts at 6 pm

tinyML Summit Program format – March 21, 2019 • Summary/Highlights of Day 1 – tinyML Technology: o HW – 20 min - Ian Bratt, ARM o Systems – 20 min, Boris Murmann, Stanford Univ. o SW – 20 min, Kurt Keutzer, UC-Berkeley • Two panels discussions moderated by Chris Rowen, BabbleLabs and Cognite Ventures o tinyML Applications: opportunities and challenges • Panelists: Edith Beigne (Facebook), Fari Assaderaghi (NXP) Ofer Dekel (Microsoft), Christoph Lang (BOSCH) o tinyML Ecosystem development • Panelists: Bill Chappell (DARPA), Jeff Henckels (Qualcomm), Zach Shelby (ARM) • Closing remarks: Call to Action – Pete Warden, Google • Lunch

tinyML Summit “Rules” • Informal • Interactive • Professional ✓ see “Code of Conduct” on the tinyML website ✓ promote diversity and comfortable, harassment-free experience ✓ respect privacy aspects

Misc “housekeeping” • Please silence cell phones • NO photos may be taken of slides • All attendees will be able to access the presentations and poster pdfs online. • Please stay for the poster/demo/networking session from 4:30 to 6:00. • The dinner will be in the café at 6:00 pm • Wifi is “Google Guest”, no password required • If you valet parked your car, please ensure you retrieve your car comfortably before 8pm on March 20, when the valet stations will close and by 3 pm on March 21 • Please stay within the designated event space and wear your badge at all times over the two days. Our event rooms have TinyML summit signs, and we have Google security staff to help us out.

Summit Dinner entrance venue Main sessions (Luca Pacioli) Restrooms Sponsors & demos Food

“Let’s make tinyML BIG !”

What is tinyML ? ? st Summit) (topic c for discussion ssion at the 1 st mit) • (for now) tinyML is broadly defined as machine learning architectures, techniques, tools and approaches capable of performing on-device analytics for a variety of sensing modalities (vision, audio, motion, environmental, human health monitoring etc.) at “ mW mW ” (or below) power range targeting predominately battery operated devices (IoT, bioelectronics, …) 16