Bias in AI

March 3, 2022 | Women In Technology

It’s not ‘AI’ ok – removing bias

When AI results cannot be extrapolated widely, bias occurs. Bias can be introduced by how data is obtained, how algorithms are designed, and how AI outputs are interpreted. Bias is generally associated with preferences or exclusions in training data, but bias can also be generated by data collection methods, algorithm design, and other factors.

Algorithmic bias in AI is a widespread issue. You may recall hearing about biased algorithm examples in the news, such as speech recognition failing to recognise the pronoun "hers" but recognising "his," or facial recognition software failing to distinguish individuals with dark skin. While eliminating prejudice in AI is unachievable, it is vital to understand not only how to lessen bias in AI, but also how to actively seek to prevent it. Knowing the training data sets that are used to develop and evolve models is key to understanding how to avoid bias in AI systems.

So how can you plan implementable actions that will help you or your team start removing bias?

Don’t try and solve everything at once

When you try to solve for too many cases, you end up with an unmanageable number of labels spread across an unmanageable number of classes. To begin, narrowly describing an issue will assist you in ensuring that your model is operating properly for the specific reason you designed it.

Get to understand your data

There are classifications and labels in both academic and commercial datasets that can introduce bias into your algorithms. You're less likely to be surprised by unwanted labels if you understand and own your data.

Structure your data collection that allows for a variety of viewpoints

For a single data point, there are frequently numerous correct opinions or labels If you collect their perspectives and allow for valid, often subjective, disputes, your model will be more adaptive.

Think about who your end-users are

Recognise that your end-users will not be identical to you or your staff. Empathise with others. Avoid AI bias by anticipating how individuals who aren't like you will engage with your technology and the issues that may occur as a result.

Include a variety of annotations

The greater the pool of human annotators, the more diverse your ideas are. This can significantly reduce bias, both during the first launch and as your models are retrained.

Incorporate feedback in model testing

Models are rarely static throughout their lives. Deploying your model without a means for end-users to provide input on how the model is performing in the real world is a common, but costly, mistake. Opening a conversation and feedback forum can help to ensure that your model continues to perform at its best for everyone.

Plan to use the input to enhance your model.

You'll want to keep reviewing your model, not just based on consumer input, but also by having independent individuals audit it for changes, edge cases, biases you could have overlooked, and so on. Make sure you collect feedback from your model and provide it with your own to help it improve its performance.

Organisations need to understand that It's preferable to uncover and address vulnerabilities immediately rather than waiting on regulators to do so later. The good news is that bias can be considerably minimised or eliminated by utilising proper modelling principles, and individuals working on AI may help uncover accepted biases, establish a more ethical understanding of hard problems, and stay on the right side of the law – whatever it turns out to be.