1 of 28

Responsible AI via Sustainable and Privacy Preserving EdgeML

Brian Plancher

Barnard College, Columbia University

brianplancher.com

1

CC BY-NC-SA 4.0

2 of 28

How can we adapt our approach to EdgeML to support Responsible AI future?

2

CC BY-NC-SA 4.0

3 of 28

3

What is a (Datasheet for a) Machine Learning Sensor?
Applications of TinyML for Sustainability
Environmental Footprint of TinyML Systems at Deployed Scale

Responsible AI via Sustainable and

Privacy Preserving EdgeML

Positive Effects of TinyML

Environmental Footprint at Scale

Our Suggested Transparent Paradigm

CC BY-NC-SA 4.0

4 of 28

What is a (Datasheet for a) Machine Learning Sensor?

4

CC BY-NC-SA 4.0

5 of 28

What is a Machine Learning Sensor?

5

Sensor 2.0

Privacy by Design

CC BY-NC-SA 4.0

6 of 28

We suggest transparency as a core value to overcome these challenges.

6

Description, Features and Use Cases

Compliance

Diagram and�Form Factor

Hardware Characteristics

Communication Specification and Pinout

Model�Characteristics

Environmental Impact

Privacy and Security

End-to-End Performance Analysis

Data�Nutrition Label

arxiv.org/abs/2306.08848

CC BY-NC-SA 4.0

7 of 28

7

Description, Features and Use Cases

Compliance

Diagram and�Form Factor

Hardware Characteristics

Communication Specification and Pinout

Model�Characteristics

Environmental Impact

Privacy and Security

End-to-End Performance Analysis

Data�Nutrition Label

ML Sensor Datasheets Have 3 Goals:

Raise the level of abstraction
Transparent at the hardware, data, model, and end-to-end layers
Support Responsible Use

CC BY-NC-SA 4.0

8 of 28

8

Description, Features and Use Cases

Compliance

Diagram and�Form Factor

Hardware Characteristics

Communication Specification and Pinout

Model�Characteristics

Environmental Impact

Privacy and Security

End-to-End Performance Analysis

Data�Nutrition Label

ML Sensor Datasheets Have 3 Goals:

Raise the level of abstraction
Transparent at the hardware, data, model, and end-to-end layers
Support Responsible Use

Standard Sensor Datasheet

CC BY-NC-SA 4.0

9 of 28

9

Description, Features and Use Cases

Compliance

Diagram and�Form Factor

Hardware Characteristics

Communication Specification and Pinout

Model�Characteristics

Environmental Impact

Privacy and Security

End-to-End Performance Analysis

Data�Nutrition Label

ML Sensor Datasheets Have 3 Goals:

Raise the level of abstraction
Transparent at the hardware, data, model, and end-to-end layers
Support Responsible Use

Responsible Machine Learning Analysis

CC BY-NC-SA 4.0

10 of 28

Responsible Machine Learning Analysis

10

CC BY-NC-SA 4.0

11 of 28

11

Description, Features and Use Cases

Compliance

Diagram and�Form Factor

Hardware Characteristics

Communication Specification and Pinout

Model�Characteristics

Environmental Impact

Privacy and Security

End-to-End Performance Analysis

Data�Nutrition Label

ML Sensor Datasheets Have 3 Goals:

Raise the level of abstraction
Transparent at the hardware, data, model, and end-to-end layers
Support Responsible Use

Overall System Analysis

CC BY-NC-SA 4.0

12 of 28

End-to-End Responsible Performance Analysis

12

CC BY-NC-SA 4.0

13 of 28

Environmental Impact

Lets Explore this Impact in More Detail

13

CC BY-NC-SA 4.0

14 of 28

14

TinyML can support the SDGs but comes with costs. What is the net impact?

CC BY-NC-SA 4.0

15 of 28

Applications of TinyML for Sustainability

15

CC BY-NC-SA 4.0

16 of 28

Zero Hunger & Good Health and Well-Being

(SDG #2 & #3)

16

Credit: PlantVillage Nuru

Nuru, an ML app more accurate than humans at detecting plant diseases. Increased a farmer’s sales by 55% & yields by 146%.

Credit: Crop Angel Ltd

Tiny drones can provide targeted pesticide applications that reduce use to 0.1% of conventional blanket spraying.

Credit: Sinhyu/Getty Images

Using Edge Impulse, a system was prototyped to identify mosquitoes by wing beats sounds with 88.3% accuracy.

CC BY-NC-SA 4.0

TLDR: ML applications can increase agriculture production through data-driven methods & can also be used to aid in our health and well-being.

Fact for last image maybe to mention: One of the diseases noted in the UN’s SDG report is

malaria due to its massive global impact. In fact, nearly half of

the whole world population has been killed by mosquitoes [50].

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

CREDITS:

From left to right:

PlantVillage Nuru: https://plantvillage.psu.edu/projects (See Nuru App)

Smart Collars: https://www.nature.com/articles/544S21a

Drones for Pesticide Spraying: https://www.nature.com/articles/544S21a

Malaria / Mosquitos: https://ieeexplore-ieee-org.ezp-prod1.hul.harvard.edu/stamp/stamp.jsp?tp=&arnumber=9662116

17 of 28

Life on Land & Below Water

(SDG #14 & #15)

17

Credit: Rainforest Connection

Rainforest Connection uses

recycled smartphones for solar-powered listening devices to warn of deforestation efforts

Credit: Tim Cole

To prevent collisions with

whales in busy waterways, Google deployed a TinyML model on hydrophones to alert ships.

Credit: RESOLVE and Bivash Pandav

RESOLVE’s AI camera transmits notifications of elephant detection and can run for more than 1.5 years on a single battery.

CC BY-NC-SA 4.0

TLDR: TinyML can help preserve the planet’s biodiversity by improving the efficiency of conservation efforts that rely on distributed sensing networks.

Note: Human-elephant conflict very common in Asia and Africa– Elephant-human conflict is a result of habitat loss and fragmentation. When elephants and humans interact, there is conflict from crop raiding, injuries and deaths to humans caused by elephants, and elephants being killed by humans for reasons other than ivory and habitat degradation

Note: Hydrophone = underwater microphones

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -

CREDITS:

Elephants Image: https://www.resolve.ngo/blog/WildEyes-AI-Helping-to-Save-Wild-Elephants-and-Prevent-Human-Elephant-Conflict.htm

Orca Image: https://venturebeat.com/ai/googles-ai-powers-real-time-orca-tracking-in-vancouver-bay/

Rainforest Connection: https://blog.huawei.com/2021/12/13/rainforest-connection-fighting-back-poachers-technology/

18 of 28

Climate Action

(SDG #13)

18

Smart HVAC systems show a 20-40% reduction in building energy usage.

Credit: Google Nest

Ribbit Network is crowdsourcing world’s largest greenhouse gas emissions dataset through distributed intelligent sensors

Credit: Ribbit Network

TinyML can help provide intelligence to tiny robots like the Robobee that can be used as artificial pollinators.

Credit: Wyss Institute at Harvard University

CC BY-NC-SA 4.0

19 of 28

Environmental Footprint of TinyML Systems at Deployed Scale

19

CC BY-NC-SA 4.0

20 of 28

Real TinyML Systems are more than just an MCU!

20

CC BY-NC-SA 4.0

21 of 28

Building Representative Systems

21

High Cost

Medium Cost

Low Cost

Image Classification

Keyword Spotting

Cost Level

Application

Size

Large

Compact

CC BY-NC-SA 4.0

22 of 28

Building Representative Systems

22

CC BY-NC-SA 4.0

23 of 28

23

harvard-edge.github.io/TinyML-Footprint/

CC BY-NC-SA 4.0

24 of 28

TinyML Systems in Context

24

5x to 38x Savings over a 3-year lifespan!

CC BY-NC-SA 4.0

25 of 28

Environmental Footprint of TinyML Systems at Deployed Scale

25

CC BY-NC-SA 4.0

26 of 28

What if we scale to 250bn devices?

26

If our devices have longer lifespans this data starts to look even better!

CC BY-NC-SA 4.0

27 of 28

Limitations and Areas for Future Study

27

What about the net impact of factors beyond carbon?

What about the

human costs?

What about

Jevons’ Paradox?

How can emerging technologies help?

CC BY-NC-SA 4.0

28 of 28

Responsible AI via Sustainable and Privacy Preserving EdgeML

Brian Plancher

Barnard College, Columbia University

brianplancher.com

28

CC BY-NC-SA 4.0