Eliminating AI bias: interview with AI Verse CEO

Electronic Specifier's Kristian McCann speaks with Benoit Morisset, CEO, AI Verse about eliminating AI bias.

"The main problem we are addressing is reducing this bias in the deep learning stage of AI systems,"says AI Verse CEO and Founder Benoit Morisset. Benoit's company is aimed at improving AI systems and reducing this bias through the datasets being fed to systems, but what is AI bias and why is it an issue for the industry?

The issue with AI bias

AI bias is when you build a system with data that doesn't illustrate a full picture of a situation. There is algorithmic AI bias or "data bias," where algorithms are trained using biased data. The other kind of bias in AI is societal AI bias, so the person who inputs data influenced by their beliefs and understandings into a system. Benoit's system is focused on improving the former.

"Let's say you want to collect data for your system, maybe data on traffic accidents involving pedestrians.You deploy a team to go to pedestrian crossings and streets and use a normal camera to photograph a million images of the different scenes and people there," says Benoit. "When they are finished, they will then upload all that data to the system. Now, with the data in the machine, you ask the AI system: how many men and how many women did you have in your images?That is something that becomes difficult because if the camera didn't apply a value to each person in the photo when it was taken, when you query the system, it cannot give you a proper answer."This method, Benoit explains, deprives the system of information it could then use to further increase its understanding and learn from the situation.

Also, what is, or isn't included in the image, could create a bias. "Now let's say when your team were out photographing, they went during summer. Now, the information you will have will likely have a over representation of sunny and bright days and less rainy, snowy or overcast days." Benoit continues:,"Or say they could only go out during weekdays from 9am to 3pm, and as a result all the kids are in school, and only working adults are out on the street. These are all ways how what you are feeding the system distort the reality of a situation." This input could then not only limit its ability to deal with situations it hasn't experienced but could lead it to make incorrect predictions or assessments of events due to this initial input it used to learn.

Controlling your datasets

AI Verse's solution to stopping AI bias is to control the input given to systems from which it will then learn. To do this, the company created an engine on the Cloud that can be given specific parameters for what the developer wants to be shown and then the system can render an image from scratch based on those specifications. "For instance, if you were developing a smart vacuum cleaner, you could train it to detect when someone is sleeping on the sofa and instead of it starting up a pre-scheduled programme, it could recognise the situation and wait for the person to wake before starting," states Benoit. "If you were trying to do this through photos sourced online it would be impossible.As first, sourcing the images, there might not be that many for it to draw from, and second, you would not be able to find as many nuanced situations of someone in that position than if you created these images yourself according to your own parameters."

Benoit believes that this quality of data is key to reducing AI bias and that this will lead to a blossoming of AI's use. "I think everything is going to merge. You can see with ChatGPT the system is making tremendous, tremendous progress." He continues:, "So now the issue you have with language, natural language understanding, is going to improve, as well as AI's understanding of what is going on in images and videos. Then you're going to have like some extremely advanced system in terms of the level of interaction you can have with them. There is going to be much more relevant, much more adapted content to the alignment of the specific conditions and everything. Everything is merging around the same type of architecture, which is the deep learning architecture."