A Summarisation of Facebook’s research paper titled “Does Object Recognition Work for Everyone?”
“AI is a tool. The choice about how it gets deployed is ours.” Oren Etzioni
AI technologies are slowly and steadily making a foray into our lives. They are already making some important decisions for us like whether we are qualified for a mortgage, what kind of movies or songs we prefer and are even suggesting email replies to us. Computer vision is one such actively growing subfield of AI which holds a lot of promise. Techniques like facial recognition, object detection, Image recognition, emotion analysis etc are being used across industries to enhance the consumer experience, reduce costs and increase security. But what if the results of these systems are prejudiced towards a particular race, gender or region. Well, there is definitely more to it than what meets the eye.
Recently, Facebook researchers in a paper titled: Does Object Recognition Work for Everyone? raised a very pertinent question regarding the applicability of the results of the current SOTA objection detection techniques. Their study highlights a strong geographical bias when it comes to recognizing common household items in low income or Non-Western countries. The figure below shows how the prediction results are far away from the ground truth in low-income nations whereas the results for Western countries are highly accurate.
Biased data leads to biased models
The story of bias and under-representation in computer vision systems isn’t new. Many pieces of research have come to light which clearly demonstrates this pattern. Here are two concrete examples.
- The paper Intersectional accuracy disparities in commercial gender classification highlight the fact that commercially used gender classification systems tend to be more accurate for lighter skinned males as compared darker females.
- Unequal representation and gender stereotypes in image search results for occupations showed that there was an under-representation of women in Google search results.
One of the most probable reasons for such biased results is the type of data used for training the model. In most of the cases, the data collection is itself biased and it is this bias that eventually percolates and shows in the model’s predictions. In the current study, it was found that the photo collections used in almost all popular datasets including ImageNet, COCO, and OpenImages. had a very skewed geographic distribution with the majority of the data from Europe and America.
It is rather surprising to see that while the majority of the population is concentrated in Africa, South America, and Central and Southeast Asia, their percentage contribution to the data is way less.
Need for a Diversified data
The study was done on the Dollar Street dataset which is a collection of the household items taken from 264 families across 50 countries. These images were gathered by the Gapminder foundation, with a view to measure the difference in the living conditions of the people living in the world.
The Dollar Street dataset, as opposed to the other popular image datasets, has a fairly distributed data. The following Cholorpeth maps show the number of images per country in the Dollar street dataset.
The Dollar street dataset also records the monthly consumption income for every house whose picture is taken. This brings to light another important factor relating to income disparity, in the current object detection systems.
“the error rate of object-recognition systems for households with an income of less than US$50 per month is approximately 10% lower compared to households making more than US$3, 500 per month”[Terrance et al.(2019)]
Here is the performance of the top models with their top 5 predictions. Pretty astounding.
The paper highlights serious issues in the data collection methodology and the impact it can have on the AI systems being developed throughout the world. In order to cultivate more trust and faith in these systems, it is essential that researches should not be limited to a particular region or economic status but should include all regions, genders and races of the society. Only then AI will be truly meaningful for all.