The role of data in machine learning

Let’s look more in-depth at the role data plays within machine learning.

Machine learning data sets

In Machine Learning, data sets are used for training and testing. A training set is a subset of data used to train a model. A test set is a subset of data used to test the trained model (1). Typically the training set will be much larger than the testing set.

A wide range of datasets and machine learning tools

There is a wide range of existing datasets and machine learning tools which are widely available such as from Google, Microsoft and IBM (2, 3, 4).

Typically, Machine Learning training sets require a large amount of data. The amount is difficult to quantify, as it depends on the complexity of the system being built but typically consider that it would require tens of thousands to millions of data points.

It can be difficult to find an existing dataset with large enough representative data for underrepresented/marginalised demographic groups.

Google

Google has taken steps to create Inclusive Machine Learning which will reduce the likelihood of negative stereotypes, automatic denial of access to services and product failure for underrepresented groups.

Their AutoML platform enables smaller training sets to be used in Machine learning increasing the inclusivity of machine learning.

Machine learning influences design

Machine learning influences design as it can impact the usability and accessibility of products. As an example, during focus groups, it may become clear that different dialects and colloquialisms need to be considered when training a voice recognition application, or chatbot.

Without using the correct training sets and particularly test sets, the system won’t be fit for purpose for these people.

Inputs and results during testing

In machine learning, the test set will typically be a subset of the training set and may represent a particular demographic. For wider types of products and services, the test data is used to represent expected usage.

These could be values used to complete a form on a website, the images of a face for an image processing algorithm, or messages entered into a chatbot.

Analysis

Analysis of the results found throughout testing can lead to a significant amount of change required when creating new products, or updating existing products.

Utilise the results to inform the design, this may mean adding accessibility options, or changing terminology so that it does not exclude cultures/demographics.

It is important to perform testing iteratively throughout development so that results can influence necessary updates to design.

References

1. Training and Test Sets: Splitting Data
2. Datasets – Google Research
3. Microsoft Azure Open Datasets
4. Artificial intelligence Datasets – IBM Developer
5. Inclusive ML – Google