7

Make your Arduino learn

TUTORIAL

Machine learning doesn’t have to require data centres: all you need is know-how and an Arduino!

Make your Arduino learn

ou’ve probably seen countless

news stories about new artificial

intelligence (AI) systems that can

learn like humans do. At the core

of these AI systems is something

called machine learning. Machine

learning is a technique that allows the internal rules of

a system to be automatically learnt using data, rather

than a programmer hard-coding in all of the rules.

There is a misconception that you always require

a lot of data and computing power to build these

machine learning systems. For example, it’s common

to hear about teams using millions of data points and

powerful servers in their machine learning efforts.

For these larger machine learning jobs, they need

those resources because they are trying to solve very

complex problems, like learning how to drive a car

around a busy city. For simpler applications such as

‘when to turn on a lamp’, the amount of resources

required can be relatively minuscule: so minuscule

that you can fit every aspect of the machine learning

process onto a single Arduino.

We’ll walk through how to build a simple ‘automatic

night light’ system (a system that measures the

ambient light in the room and turns on an LED, based

on some defined threshold) with a machine learning

twist; rather than hard-coding the ambient light

threshold for which to turn on the LED, the machine

learning system will learn when to turn on the LED

based on data collected regarding when the user has

manually turned the LED on or off.

You will start off by toggling the LED on or off, as

you normally would by simply using a button, and then

the system will begin to learn and perform the task

automatically. What’s especially powerful about this

system is that it can be given to many different people,

and it will learn each of their individual preferences

without them changing code or fiddling with a settings

knob, as in traditional night light systems. As well

as the system’s adaptability to different people, as

your own preferences change over time, so too will

the system’s. Behold, this is the power of machine

learning systems.

There are three main processes of machine learning

systems: data collection, training/learning, and

prediction. Larger-scale applications would normally

have these different tasks done separately (e.g. collect

and store user data using a web form and database,

then use servers to perform model training, then

finally save the trained model into a web application

that can be asked to make predictions on new data).

Our system will have all of the same processes, but

running on just one Arduino. Although these processes

run in the order: 1) collect data, 2) train model, 3) make

predictions, it is easier to explain the development of

these processes in reverse order.

PREDICTION

In order to solve a problem with machine learning, you

need to ‘phrase’ the problem with a set of variables

that need to be optimised. For example, in the real

world you might want to boil an egg perfectly – to do

this you need to find the optimal heat level and cooking

time. So, if we were using machine learning we would

need a variable for ‘heat_level’ and ‘cooking_time’,

and it is these variables that will get automatically

Y

Zack Akil

Zack is a data scientist

focusing on the big

problems, such as the

proper filming of

rugby matches.

ZackAkil

Below

Our final circuit

lighting up our

dark evenings