Scribd
Upload
9 views7 pages

Introduction To AI - Part Three

Uploaded by

Tgowtham Arasu
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
9 views7 pages

Introduction To AI - Part Three

Uploaded by

Tgowtham Arasu
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 7

Understand the basics

and the history of AI.


This section unpacks the important parts of a ML
model from the supervised or unsupervised learning
families. Reinforcement learning algorithms work in
a slightly different framework which we won’t cover
here.

Figure 3: A machine learning model consists of an algorithm,


together with some parameters. The parameters are learned from
the training data. To test the ability of the model to generalise it
must be applied to new data, that wasn’t used for training.
A machine learning model consists of an algorithm,
together with some parameters.

The algorithm consists of a


series of computational
steps or instructions on how
to manipulate the incoming
data to achieve the chosen
task.

The parameters are values


that are used by the
algorithm at different stages.

Example
An algorithmic step might say to multiply the input
data by a scaling factor, where the particular value
of the scaling factor is a parameter learned from the
data during training.
In ML applications, the algorithm is usually fixed while
the parameters are learned from the data in a process
called training. It is important to understand that the
performance of a model is function of both the
algorithm and the parameters, and the parameters
come from the data used during training.

Usually, an ML model is developed so that it can be


applied to new data. If model performance (i.e. for the
face detection classifier, the proportion of correct
answers, called the accuracy) is evaluated on the
training data it is usually much higher than the
performance would be on new data.
Because it is being asked about a photograph it has
already seen in training, it can know the correct
answer for this specific example directly, for example
by memorising every photo in the training set, without
having learned a general process to solve the task
(i.e. recognising faces).

The ability of a model to work successfully on new


data is called generalisation. We want models that
generalise to new data, and not models that just
memorise the training data. The phenomena where
models perform much better when evaluated on their
training data is called over-fitting.
To meaningfully evaluate model
performance, it is very important that
the data used to test performance is
completely separate from the data
used to train the model. When new
data is not immediately available this
is achieved with techniques like
cross-validation.

In cross-validation, the available


data is randomly split into two parts,
one part is used to train the model
and the second part is used to test
the model. This allows researchers to
approximate the performance of the
model on new data.
You can find the
full guide and other
resources at our
website.

You might also like