Thanks to advanced algorithmic innovations over the past years, computers have started to ‘make sense’ of the world more vividly. You’ll be surprised to realise that despite these innovations, even the most upgraded Artificial Intelligence (AI) vision system will interpret your face with racist slurs, gender stereotypes or any other term causing disrepute.
Scientists who were involved in Machine Learning (ML) for computers are now removing some of the human prejudice in the data used to teach algorithms. The effort indicates removing bias from AI systems is quite tedious because it is still heavily reliant on human training. Olga Russakovsky, Assistant Professor at Princeton University, said that the process requires re-considering “a lot of things”.
It is a crucial issue because the rapid deployment of AI systems across the world and in various sectors can pose significant real-world consequences. Instances of bias have already been identified in facial recognition systems, hiring programmes and web-searching algorithms. A horrifying scenario involves AI vision systems misidentifying minorities before police as ‘criminals’.
In 2012, Project ‘ImageNet‘ played a critical role in unlocking AI’s potential by allowing developers to teach visual concepts to computers through ML including objects like ‘flowers’. Scientstists from Stanford, Princeton and the University of North Carolina paid Amazon Mechanical Turks (MTurks) a fee to label more than 14 million images.
When this dataset was fed into a vast neural network for image-recognition purposes, it was able to identify objects with surprising accuracy. ML enabled the AI algorithms to learn from examples to for pattern identification such as pixels which make up the shape and texture of puppies. A contest was also launched for testing algorithms which correctly classified humans. The success of ImageNet systems triggered excitement for AI-enabled image recognition and ushered in technologies such as advanced smartphone cameras and automated vehicles.
Since those achievements, some other researchers found problems within ImageNet data. Its algorithms could assume that programmers are white-coloured men because images on which the AI system was trained labelled “programmer” as such. Recently, Excavating.AI also highlighted human prejudice in labelling for ImageNet such as “radiologist” and “puppeteer” as also racial slurs like “negro” and “gook”. The now-defunct website allowed people to submit photos and see terms in the AI model trained using the dataset of ImageNet. These were probably caused by prejudicial and biased labelling by MTurks.
ImageNet’s team identified and acknowledged biasness in their dataset and initiated steps to remove them. Crowdsourcing was employed to identify and remove derogatory words. In addition, they also identified terms projecting meaning on images suchas “philanthropist” and recommended their exclusion from the ML process.
Furthermore, the team at ImageNet assessed demographic and geographic diversity in its dataset. Previously, “programmer” would revealphotos of white men but now it shows greater diversity to re-train AI algorithms.
The whole process simply shows that AI can be re-configured to produce fairer results while also revealing that extent to which machines rely on human knowledge for training.
Andrei Barbu, Research Scientist at MIT who studied ImageNet, admired the effort. But he did caution that identifying only a few problems would remove only little of the inherest bias to balance things out properly. He says that stripping out bias could make the dataset “less useful”, especially when race, gender and age are taken into account. He also says, “Creating a data set that lacks certain biases very quickly slices up your data into such small pieces that hardly anything is left“.
Russakovsky agrees that the remedial process is complex but believes it would fare well in the long term. She said, “I am optimistic that automated decision making will become fairer…Debiasing humans is harder than debiasing AI systems“.