AI Virtual Tech Talks Series Demystify artificial intelligence on Arm MCUs Cartesiam
Joël Rubino (CEO) & Francois de Rochebouët (CTO) July 2020 CARTESIAM WE HELP EMBEDDED DEVELOPERS MAKE OBJECTS SELF-AWARE
2 https://discord.gg/rcTyQrY AGENDA
Joël Rubino (CEO) • Who we are / The problem we solve 10
François de Rochebouet (CTO) • Live creation of a « smart sensor » 35
Q & A
15
3 https://discord.gg/rcTyQrY FRANCE Toulon HQ - R&D
Paris SALES & MARKETING
Software company founded in 2016 USA #DNA Focus on Machine Learning at the Edge New-York NORTH AMERICAN OPERATIONS
#MISSION To help developers to easily and quickly create smart objects without AI skills
#INNOVATION Training and inference exactly where data turn into a signal GERMANY Munich Strong scientific and business background, SALES #TEAM PHDs in advanced mathematics, machine learning, signal processing
4 https://discord.gg/rcTyQrY « Edge AI » is a catch-all word. Training and Inference are different
Train the model on GPU/CPU Training Cloud
Apply the model « at the Edge » Inference Device
5 https://discord.gg/rcTyQrY « Edge AI » is a catch-all word. CURRENT AI APPROACH IS IRRELEVANT FOR OBJECTS
Train the model on GPU/CPU Training Cloud
Data set
Data scientists
EXPENSIVE, YEARS-LONG and COMPLEX Projects NEVER GO FURTHER THAN A POC
https://www.gartner.com/smarterwithgartner/3-barriers-to-ai-adoption/ 6 https://discord.gg/rcTyQrY « Edge AI » is a catch-all word. Training and Inference are different
OPTION 1
Make this easier, simpler, faster for Train the model on GPU/CPU Training embedded Cloud dev / data- scientists
Apply the model « at the Edge » Inference Device
7 https://discord.gg/rcTyQrY « Edge AI » is a catch-all word. A datascientist in an MCU
OPTION 2
IDE for embedded dev to create the best ML library for each project Win / Linux / Win
Train the model « at the Edge » • Dynamic learning “on device” Training + Inference • Any embedded developer at the Edge Anomaly detection • Any arm Cortex M0 - M7
8 https://discord.gg/rcTyQrY HOW ?
NanoEdge™ AI Studio : REAL EDGE MACHINE LEARNING
We re-wrote (starting from the algebra) every single ML We created NANOEDGE AI studio, a and signal processing algorithms so they can run software for embedded developers, inside a microcontroller (learning and inference) running autonomously on a PC, and capable of simply generating the right • Ultra optimized / customized for optimal performance learning machine library for your project • What is not used is not loaded from among 500 million possibilities. • Ultra memory efficient • Super specialized models • Learning and analysis inside the microcontroller
9 https://discord.gg/rcTyQrY Expand ML usage below and beyond the obvious sweet spot
Cortex-M Today New Cortex-M (enabled by Helium)
Cortex-M7 Well suited for ML and Cortex-M55 DSP application Cortex-M35P Cortex-M33 Cortex-M3 Cortex-M4
Cortex-M23 Cortex-M0+
Relative control code performance code control Relative Cortex-M0 Cortex-M1
Relative ML and signal
10 https://discord.gg/rcTyQrY Expand ML usage below and beyond the obvious sweet spot
Cortex-M Today New Cortex-M (enabled by Helium)
Cortex-M7
Cortex-M55 Well suited for ML technology Cortex-M35P from Cartesiam Cortex-M33 Cortex-M3 Cortex-M4
Cortex-M23 Cortex-M0+
Relative control code performance code control Relative Cortex-M0 Cortex-M1
Relative ML and signal
11 https://discord.gg/rcTyQrY WE ARE THE AI POSTER CHILD FOR THE MICROCONTROLLER LEADERS…. AND IT GOT NOTICED !
PARTNERSHIP ANALYSTS
2019 “Hype Cycle for Semiconductors and 2019 “Enterprise of the year” Electronics Technologies TOP 3 EDGE AI Category ’AI at the Edge’ VENDORS “ AI Partner 2020 “Cartesiam named “Cool Vendor 2020 IOT”
TESTIMONIAL INDUSTRY
2019 WINNER 2019 WINNER Category ‘Technology Category ‘Most Innovative Showcase of the year’ Technology’
CUSTOMERS
…
12 https://discord.gg/rcTyQrY CURRENTINDUSTRY AIAND APPROACH USE-CASE AGNOSTICIS THREE CLIENT EXAMPLES OUT OF ENDLESS POSSIBILITIES IRRELEVANT FOR OBJECTS
Predictive maintenance Home appliance upgrade Confidential Endless use cases Industry Mass market Defense Any Industry
Bob Assistant™ learns the In Japan, a home appliance French Navy is currently patterns of any engine and leader upgraded its offerings using our solution. sends alerts if anomalies to provide filter clogging We have no clue about the occur detection on air conditioning use case as all happens on Thousands Units sold across systems using current the PC with no data sent to Europe monitoring. Cartesiam/cloud.
Software update on existing hardware. Time to market massively reduced. Works with all compatible filters. 13 https://discord.gg/rcTyQrY CURRENTLY AVAILABLE SIGNALS
We can handle : All Cortex M Microcontroller • Vibrations • Pressure • (VOC) Volatile organic compound • Current • Magnetic • Temperature • Ultrasound • Multi signal sensor
We currently do not handle: • Video • Voice
14 https://discord.gg/rcTyQrY ONE SOFTWARE TO DELIVER 500 MILLION MACHINE LEARNING LIBRARIES
DEMO
15 https://discord.gg/rcTyQrY AI Virtual Tech Talks Series Thank You Danke Merci 谢谢 ありがとう Gracias Kiitos 감사합니다 ध यवाद ﺷﻛ ًرا ות הדו