11

Thinking Pi

TUTORIAL

Transform your Raspberry Pi into a silicon brain using scikit-learn

Thinking Pi

here’s a lot of talk flying around

about super-powerful artificial

intelligence (AI) running on GPU

clusters in the cloud. This makes

AI sound like it’s just for people with

their own supercomputers and PhDs.

However, you can easily build your own learning

machines on practically any hardware you want.

In fact, one of the top machine learning libraries is

a simple Python library that can be running on your

Raspberry Pi in minutes!

AI can be a bit of a troublesome phrase, not least

because no one seems to agree on what intelligence

actually is. We’re going to focus on one aspect of

intelligence: learning. This is the ability to do actions

based on what you’ve observed in the past – or in this

case, your Raspberry Pi performing actions based on

what input it’s previously received.

Building and running machine learning on devices

like Raspberry Pis allows you to build clever machines

using whatever data you want (public or private), and

apply that to applications completely untethered from

the internet. A common example is a thermostat that

not only lets you control it with a quirky interface, but

also learns what temperature to set, based on time

of day, humidity, who is in the room, and learnt solely

from how you have set the temperature in the past.

Getting into machine learning can seem like a

huge undertaking if you have never touched it before.

Fortunately, one of the most popular machine learning

libraries is founded on the principle of being as simple

as possible to use. Scikit-learn is an open-source

comprehensive library that has implementations of the

most common machine learning algorithms, packaged

in a user-friendly API. Let’s install it in onto our

Raspberry Pi using pip, (a package installer for Python)!

pip install -U scikit-learn

Before we jump into the library, here is a little

machine learning intuition.

In the type of machine learning we are going to do

(supervised learning), our code is going to learn the

pattern that maps a set of inputs to a set of outputs.

What does this mean? If you were a botanist (plant

expert) who sees a bunch of different flowers that

you don’t know the name of, and another expert

points to a couple of unique flowers and tells you

their name, in your head you are learning a mapping

between the unique features of the flowers (e.g.

colour, shape, smell) and the name of the flower.

So, when you see another flower, you can use that

‘mapping’ to predict the name of the flower, based on

its features, without the other expert’s help. So the

flowers’ features (e.g. colour, shape, smell) are the

inputs, and the names of the flowers are the outputs.

Our machine learning algorithm will be tasked with

learning the mapping between them, so that with

just inputs (flower features), it can predict the outputs

(flower name).

SIGNAL AND NOISE

The most important skill in machine learning is in

developing a good intuition around ‘how much’ and

‘what kind’ of data you need to learn a mapping/

pattern. Imagine you have no idea about the rules of

Quidditch (maybe you don’t have to imagine), and

your task is to figure out the rules just by watching it

being played. How many matches would it take you?

How many matches go to penalty shoot-outs? If you

don’t see it, you won’t know that it’s part of the rules,

so you’ll have to see enough to see how every rule

applies to every scenario. What happens if, in some of

the matches, the referee is making wrong calls? You

will learn the wrong rules, unless you watch enough

matches where the referee is making the correct calls

to compensate.

T

Right

Scikit-learn is a

hugely powerful

library that we’ll only

just be scratching the

surface of. There are

many more features

if you dig beneath

the surface

Zack Akil

Zack is a data scientist
focusing on the big
problems, such as the
proper filming of
rugby matches.

ZackAkil

YOU’LL NEED

Raspberry Pi