11

Thinking Pi

FORGE

In the example of learning the rules of Quidditch

based on pure observation, you can imagine that the

simpler the rules, the fewer observations you would

need in order to learn the rules. The same principle

applies to all machine learning: the more complex the

problem, the more data you will need to learn it, and

vice versa. The other issue of compensating the anti-

learning effect of seeing a bad referee by seeing even

more good referees, can be mirrored within machine

learning as well. It’s known as a noisy data problem

(signal-to-noise for Nate Silver fans). The noisier the

data (i.e. the more examples that don’t perfectly

match the pattern you are trying the learn), the more

data your machine learning model is going to need to

compensate for it.

HELLO PETAL

It should come as no surprise that the problem we

will be solving using the Raspberry Pi is a simple one

with very clean data (and because of that, we won’t

need very much data). The problem is predicting what

type of flower we are looking at, based on different

attributes (or ‘features’ in machine learning lingo) of

the flowers, e.g. petal length and width. First, let’s get

the data!

If you haven’t done interactive Python coding before,

you’re in for a treat! In the Raspberry Pi command line,

you open an interactive Python session by typing in

and running (pressing ENTER) the following:

python

All commands will now run in Python as they

are entered.

Let’s import our flower data from scikit-learn.

It comes with nice toy datasets to play with.

from sklearn.datasets import load_iris

iris = load_iris()

iris.keys() shows all of the different things stored

inside the iris object:

iris.keys()

>>> [‘target_names’, ‘data’, ‘target’, ‘DESCR’,

‘feature_names’]

iris[‘data’] contains the input data (flower

features) for 150 different flowers in a matrix, where

each row is a different flower, and each column is a

different feature.

Pattern complexity

Amount of data needed

Noise in data

Left

Noisy data and

complex problems

require a lot of

data to learn

Below

The first five rows

of flower data