Democratization of Artificial Intelligence (AI) to Small Scale

Farmers: A Framework to Deploy AI Models to Tiny IoT Edges That

Operate in Constrained Environments

Chandrasekar Vuppalapati

, Anitha Ilapakurti

, Sharat Kedari

, Jaya Vuppalapati

Santosh Kedari

and Raja Vuppalapati

Hanumayamma Innovations and Technologies, Inc., 628 Crescent, Fremont, U.S.A.

Hanumayamma Innovations and Technologies Private Limited, HIG-II, Hyderabad, India

Keywords: Edge, IoT Device, Artificial Intelligence, Kalman Filter, Dairy Cloud, Small Scale Farmers, Hardware

Constrained Model, Hanumayamma, Cow Necklace.

Abstract: Big Data surrounds us. Every minute, our smartphone collects huge amount of data from geolocations to next

clickable item on the ecommerce site. Data has become one of the most important commodities for the

individuals and companies. Nevertheless, this data revolution has not touched every economic sector,

especially rural economies, e.g., small farmers have been largely passed over the data revolution, in the

developing countries due to infrastructure and compute constrained environments. Not only this is a huge

missed opportunity for the big data companies, it is one of the significant obstacle in the path towards

sustainable food and a huge inhibitor closing economic disparities. The purpose of the paper is to develop a

framework to deploy artificial intelligence models in constrained compute environments that enable remote

rural areas and small farmers to join the data revolution and start contribution to the digital economy and

empowers the world through the data to create a sustainable food for our collective future.

1 INTRODUCTION

Artificial intelligence (AI) stands out as a

transformational technology of our digital age - and

its practical application throughout the economy is

growing apace (Chael et al., 2018). One of the chief

reasons why AI applications are getting prominence

and industry acceptance is in its software ability to

learn, albeit continuously, from real-world use and

experience, and its capability to improve its

performance(Chael et al., 2018). It is no wonder that

the applications of AI span from complex high-

technology equipment manufacturing to personalized

exclusive recommendations.

Nevertheless, this data has not touched every

economic sector, especially rural economies, e.g.,

small farmers have been largely passed over the

revolution, in the developing countries due to

infrastructure and compute constrained environments

even when AI is critical for food sustainability (Luiz,

2019).

https://orcid.org/0000-0003-2261-759X

In noting the promise and challenge of AI, the

McKinsey Global consulting Firm noted numerous

use cases across many domains where AI could be

applied and for these AI-enabled interventions to be

effectively applied, several barriers must be

overcome (James & Jacques, 2018). These include

the challenges of data, computing, and talent

availability, as well as more basic challenges of

access, infrastructure, and financial resources that are

particularly acute in remote or economically

challenged places and communities.

One chief reason, importantly, for AI not touched

every economic sector is the current AI algorithms

are only made to run on very powerful research

workstations without considering how they can be

used on real-world hardware, embedded constrained

hardware. Machine learning in embedded systems

specifically target embedded system to gather

data, process data, and apply mathematical rules to

produce insights (Van, 2019). The embedded systems

typically consists of low memory, low Ram, limited

power compared to regular computers. Increase in

652

Vuppalapati, C., Ilapakurti, A., Kedari, S., Vuppalapati, J., Kedari, S. and Vuppalapati, R.

Democratization of Artiﬁcial Intelligence (AI) to Small Scale Farmers: A Framework to Deploy AI Models to Tiny IoT Edges That Operate in Constrained Environments.

DOI: 10.5220/0009358706520657

In Proceedings of the 9th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2020), pages 652-657

ISBN: 978-989-758-397-1; ISSN: 2184-4313

factors such as processing power (Devin, 2016) leads

to higher accuracies – the cost to bear is battery life.

To successfully disseminate AI to masses and

enable successful democratization, we need to bring

rural communities to digital revolution (people,

technology and data together) and the purpose of the

paper is to achieve such digital revolution. The paper

proposes AI deployment to Small Foot Print (Tiny)

IoT Edge device that operate in constrained device

and presents the data collected in real production

environment. Our goal is the achieve AI for all, a true

Fourth Industrial Revolution (ANDREW, 2019).

2 DEMOCRATIZATION OF AI

The democratization of AI is need of the day. The

current AI is more applicable for businesses and end-

consumers who are mostly city dwellers. Lack of data

that could potentially help local businesses and

societies are one of the most significant challenges in

AI adoption. One limiting challenge is the availability

of Data for social good use cases, especially in rural

communities. Lack of technologies, importantly, in

the hands of the users exacerbating the issue. For

instance, global penetration of Internet in rural areas

are very low compared to suburban and urban area

and this is persisting wider digital gap between rural

and urban area (ANDREW, 2019). Lack of internet

connectivity is causing inhibition of digital data

services dissemination, resulting into sparse vital

datasets capture for better governance purposes and

preventing rural population to participate in the

digital economy.

2.1 Waves of Compute

AI applications can be categorized into four waves of

Compute / AI (please see Fig 1. and Table 1) (Neil

and Michael, 2019) and with the 5G, the invent of

fifth wave is on horizon (Kai-Fu, 2018).

Figure 1: Waves of Computing.

Table 1: Waves of Compute & AI.

First Wave of

compute -

Mainframe

Characterized by processing of

Large Scale Social Media Analytics.

Use Cases include: click streaming,

personalization, exclusive

recommendations and notifications

to en

-user.

Second Wave of

compute –

client-server

Business AI developed during the

Second Wave. The algorithms were

trained on proprietary business

datasets and enabled business

analytics.

Third Wave of

compute –

Mobile

“Perception A.I.”— gets an upgrade

with eyes, ears, and myriad other

senses, collecting new data that was

never before captured, and using it

to create new a

lications

Fourth Wave

compute – Core

IoT Ed

Autonomous AI – start of

autonomous cars. IoT fuelled the

fourth wave.

Fifth Wave

(Simon, 2019)

5G Infused AI – the 5G network

provides speeds that AI requires.

2.2 Data

Small Farmers make big contribution to agriculture

and dairy production in developing countries (Devii,

2017). Unlike the dairy farms of the west, milk

originates in highly decentralized villages with the

help of small farmers who own three to five cattle and

they bring milk twice a day to milk collection centers

to get paid (Prahalad and Stuart, 1999). Simply put,

the livelihood of roughly 2 billion people (26.7% of

the world population) of small farmers in developing

world depend on agriculture and the climate change

adversely impacting their survival (see fig 2) (Robert,

2009).

Figure 2: Farmer Protest

Additionally, lack of data for serving these

farmers putting food sustainability and food security

in a huge risk mode.

Democratization of Artiﬁcial Intelligence (AI) to Small Scale Farmers: A Framework to Deploy AI Models to Tiny IoT Edges That Operate

in Constrained Environments

653

2.3 The Last Mile – Constrained

Compute Environments & “Ai

Chasm”

Fourth wave of compute has spurred the development

of Edge devices. Edge devices come in various forms

and shapes with varying compute capacities (Class 0,

Class 1 and Class2) (Bormann, Ersue and Keranen,

2014). Class 0 and Class 1 devices collect vast

amount of environmental & geolocation data on a

periodic basis. Due to constrained environments, the

class 0 devices require external devices such as

gateways & mobile phones to relay to the Internet.

However, these devices deployments for small farmer

is sporadic.

2.3.1 Class 0 Devices

Class 0 devices are very constrained sensor-like

motes. These devices are so severely constrained in

memory and processing capabilities that most likely

they will not have the resources required to

communicate directly with the Internet in a secure

manner.

In order to connect Class 0 devices to the Internet,

larger compute devices such as desktop workstations

or central nodes acting as proxies, gateways, or

servers are required at the site of Class 0 devices

deployment. For device management purposes, class

0 device could answer keep alive signals and send on/

off or basic health indications.

2.3.2 Class 1 Devices

Class 1 devices are quite constrained in code

execution space (Stack & Register Level) and

processing capabilities, such that they cannot easily

talk to other Internet nodes employing a full protocol

stack such as using Hyper Text Transport Protocol

(HTTP), Transport Layer Security (TLS), and related

security protocols and Extensible Markup Language

(XML)-based data representations(Bormann, Ersue

and Keranen, 2014). However, Class 1 devices are

capable enough to use a (Internet Protocol (IP) stack

specifically designed for constrained nodes (such as

the Constrained Application Protocol (CoAP) over

UDP [COAP]) and participate in meaningful

conversations without the help of a gateway node.

Therefore, they can be integrated as fully developed

peers into an IP network, but they need to be

parsimonious with state memory, code space, and

often power expenditure for protocol and application

usage (Ilapakurti et.al, 2017).

2.3.3 Class 2 Devices

Class 2 devices are less constrained and

fundamentally capable of supporting most of the

same protocol stacks as used on notebooks or servers.

However, even these devices can benefit from

lightweight and energy-efficient protocols and from

consuming less bandwidth. Examples of the devices

include Smartphones.

2.3.4 Constrained Device (Tiny IoT Edge)

Architecture

As shown in Fig.3, constrained devices are limited by

the compute power, memory, storage space, stack

space and work in limited infrastructure. In general,

the devices have a central microcontroller as a

processing unit with sensors tied to the device unit.

The Sensors collect data on time frequency –

frequencies of collection would affect the battery

useful time.

Figure 3: Constrained Device Architecture.

2.4 Hardware in Constrained

Environment

Small Tiny Edge devices or Class By addressing

following challenges by private sector (see Fig 4),

purpose built hardware manufacturers, AI

developers, Cloud providers and local & national

governments, we can achieve AI for all, a true Fourth

Industrial Revolution (Sanjeev, 2018):

 Infrastructure Conditions

 Operating Environment

 Device Characteristics

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

654

Figure 4: Purpose Built Hardware vs. Constrained

Environment.

3 ML MODEL FRAMEWORK

We have deployed model as part of Cow Necklace.

The following hardware consists of Accelerometers,

Gyroscope, Temperature, Humidity and on-board

Bluetooth connectivity.

The Sensor module is built (see Fig 5) on working in

constrained environments (Kedari et al, 2017).

Figure 5: Mobile App.

The Cow Necklace sensor connects to mobile

(See Fig. 6) using Bluetooth Low Energy (BLE) and

uploads data to the Dairy Analytics Cloud.

The Machine Learning Model in constrained

environment is subjected to various constraints and

trade-offs (Jiawei et al, 2011).

• Hardware to Model Accuracy

• Model Accuracy to Connectivity

• Connectivity to Hardware

Figure 6: Cow Necklace (Tiny IoT Edge) - PCB Board.

The Sensor collected data (see Figure 7):

Figure 7: Sensor Data.

Figure 8: Hardware-ML Model-Connectivity Framework.

The balance has to be drawn with respect to

applicability vs. model accuracy (see Fig 8). For

instance, if ML deployed model is active learner (e.g.,

K-means Cluster), the power consumption is taxed

very high as the algorithm dynamically allocates K

values. On the other hand, if deployed model is Lazy

learners, the model evaluation is based on the

memory resident stack space algorithm evaluation.

Here, the rules are in high-drive mode to execute the

model.

Democratization of Artiﬁcial Intelligence (AI) to Small Scale Farmers: A Framework to Deploy AI Models to Tiny IoT Edges That Operate

in Constrained Environments

655

3.1 Hardware-model Accuracy Table

(Constraint – Connectivity)

For evaluating various conditions that are subjected

to Hardware to Model Accuracy, holding

Connectivity, an infrastructure aspect, as a constraint

the following to be considered:

Model: Hardware vs. Model Accuracy

Constraint: Connectivity

The connectivity, which is infrastructure service,

could vary based on the geographical location:

Connectivity Options:

• Wi-Fi

• Manual (Bluetooth Low Energy)

• No Connectivity

3.1.1 Constraint – Wi-Fi Connectivity

With Wi-Fi availability, the model could be updated

during the hardware refresh or via over the air (OTA)

(See Fig 9). With OTA, would provide more

flexibility as the latest model could be deployed.

Let us run through the different options:

Figure 9: Connectivity: Manual.

Over the Air Model Update

This option provides more flexibility as it has

influence on the model in-memory and storage

options (see Fig 10):

Hardware – Memory: Low

• The most optimized & updated models could be

deployed on the Sensor

Hardware – Power: High

• Since on-board Wi-Fi consumes considerable

power

Hardware – Storage: Low

• Sensor collected data is posted to backend server

on a periodic basis

3.1.2 Constraint – Manual Connectivity

With manual connectivity, either Bluetooth Low

Energy, the Model execution and Hardware have

huge performance or tax penalties. Let us look

following cases:

Figure 10: Connectivity: Wi-Fi.

In this case, no OTA applicable as sensor is not

connected directly to the Internet. For Model update

during Hardware refresh, following are considered:

Hardware – Memory: High

High due to self-contained model with high

memory - host models (for historical & Outlier

detection)

Hardware – Battery: High

• High – to support in-memory & compute

operations

Hardware – Battery: Storage

• Since no connectivity, the data collected to be

saved on

Hardware design consideration: Toggle of Sensor

ambient indicators (LEDs or Speaker) provide visual

clues & delivers insights.

3.1.3 Kalman Model Code

The following code predicts Kalman Temperature

(Rajaraman and Ullman, 2011):

# Formulas

def

TempPrediction(PreviousEstimate,currentMeasur

ement,PreviousErrorInEstimate):

ErrorInEstimate = 2

ErrorInMeasurement= 4

KalmanGain = ErrorInEstimate

/(ErrorInEstimate + ErrorInMeasurement)

CurrentEstimate = PreviousEstimate +

KalmanGain*(currentMeasurement -

PreviousEstimate)

# step2

ErrorInEstimate = (1-

KalmanGain)*(PreviousErrorInEstimate)

return CurrentEstimate,ErrorInEstimate

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

656

4 CONCLUSIONS

Democratization of artificial intelligence is the need

of the day. It is our responsibility to develop models

and hardware equipment that enable the collection of

the data from the constrained environments so as to

model the AI for food sustainability and threats that

we face as humans – climate change. Finally, it is our

ardent believe that the data is our best defense and

the savior against the negative effects of climate

change. The sooner we embark on democratization

of AI to small farmers, the better we leave our

progeny a wonderful life on the earth, i.e., better than

what we have inherited.

ACKNOWLEDGEMENTS

We are very thankful to the management of

Hanumayamma Innovations and Technologies, Inc.,

for providing Sensor and Sensor data to publish as

part of the paper.

REFERENCES

Chael Chui, James Manyika, Mehdi Miremadi, Nicolaus

Henke, Rita Chung, Pieter Nel, and Sankalp Malhotra,

2018, Notes from the AI frontier: Applications and

value of deep learning, https://www.mckinsey.com/

featured-insights/artificial-intelligence/notes-from-the-

ai-frontier-applications-and-value-of-deep-learning .

Luiz Amaral - March 05, 2019, "The Data Revolution

Hasn't Yet Hit Agriculture", https://www.wri.org/blog/

2019/03/data-revolution-hasnt-yet-hit-agriculture

James Manyika and Jacques Bughin, OCT 2018, the

promise and challenge of the age of artificial

intelligence, https://www.mckinsey.com/featured-

insights/artificial-intelligence/the-promise-and-

challenge-of-the-age-of-artificial-intelligence

Van Ranst, W. Real World Applications of Artificial

Intelligence on Constrained Hardware. 2019. Web.

ANDREW PERRIN, May 2019, Digital gap between rural

and nonrural America persists, https://www.

pewresearch.org/fact-tank/2019/05/31/digital-gap-bet-

ween-rural-and-nonrural-america-persists/

Neil Sahota and Michael Ashley, Own the A.I. Revolution:

Unlock Your Artificial Intelligence Strategy to Disrupt

978-1260458374

Kai-Fu Lee, October 22, 2018, The Four Waves of

A.I.,https://fortune.com/2018/10/22/artificial-

intelligence-ai-deep-learning-kai-fu-lee/

Devin Coldewey, November 21, 2016, “Overclocked

smartwatch sensor uses vibrations to sense gestures,

objects and locations”, https://techcrunch.com/

2016/11/21/overclocked-smartwatch-sensor-uses-

vibrations-to-sense-gestures-objects-and-locations/

Simon Segars, “The Fifth Wave of Computing Is Built on

AI, 5G and a Secure IoT”, April 17, 2019 , URL:

https://www.arm.com/blogs/blueprint/the-fifth-wave-

of-computing-ai-5g-iot, Access Date: Sep 18, 2019

Devii R. Rao, “Small Farmers Make Big Contributions:

Dairy Production in India”, Published on: February 7,

2017, URL: https://ucanr.edu/blogs/blogcore/

postdetail.cfm?postnum=23184

Robert T. Watson, “Industrial Agriculture and Small-scale

Farming”, Publish Date: 2009, URL:

https://www.globalagriculture.org/report-topics/

industrial-agriculture-and-small-scale-farming.html

Bormann, Ersue and Keranen, May 2014, “Terminology for

Constrained-Node Networks”, URL: https://tools.

ietf.org/html/rfc7228

Sanjeev Sharma, Nov 28, 2018, How AI can increase the

reach of India's public welfare programmes and make

them efficient, https://economictimes.indiatimes.com/

tech/software/how-ai-can-increase-the-reach-of-indias-

public-welfare-programmes-and-make-them-

efficient/articleshow/66841527.cms?from=mdr

C. K. Prahalad, Stuart L. Hart, 1999, Strategies for the

Bottom of the Pyramid: Creating Sustainable

Development, http://pdf.wri.org/2001summit_hart

article.pdf

Anitha Ilapakurti, Jaya Shankar Vuppalapati, Santosh

Kedari, Sharat Kedari, Chitanshu Chauhan,

Chandrasekar Vuppalapati. "iDispenser — Big Data

Enabled Intelligent Dispenser", 2017 IEEE Third

International Conference on Big Data Computing

Service and Applications (BigDataService), 2017

Santosh Kedari, Jaya Shankar Vuppalapati, Anitha

Ialapakurti, Sharat Kedari, Rajasekar Vuppalapati,

Chandrasekar Vuppalapati. "Chapter 99 Adaptive Edge

Analytics - A Framework to Improve Performance and

Prognostics Capabilities for Dairy IoT Sensor",

Springer Nature, 2018

Jiawei Han, Micheline Kamber and Jian Pei, Data Mining:

Concepts and Techniques, Morgan Kaufmann; 3

edition (July 6, 2011)

Anand Rajaraman and Jeffrey David Ullman, Mining of

Massive Datasets, Cambridge University Press

(December 30, 2011)

Democratization of Artiﬁcial Intelligence (AI) to Small Scale Farmers: A Framework to Deploy AI Models to Tiny IoT Edges That Operate

in Constrained Environments

657