A new study from Facebook AI Research evaluates common machine-learning classifiers' ability to label photos of objects found in households in rich countries versus household objects from poor countries and finds that the models' performance lags significantly when being asked to classify the possessions of poor people.
Partly that's due to differences in the objects themselves (poor countries' popular brands of dish soap are simply unrecognized by the system and people in poor countries may use products or categories of products that are unknown in rich countries) and partly it's due to the circumstantial differences of people living in poor countries -- for example, the classifiers struggled to recognize toothbrushes when they were appeared outside of the bathroom, as may be the case in people who live in a single room.
It's a fascinating look at the thorny problem of sampling and training bias. More interestingly, it's a great mcguffin for a techno-thriller, in which crooks buy up poor-world objects to confound security systems' classifiers.
More importantly, our study has identified geographical and income-related accuracy disparities but it has not solved them. Our analysis in Section 3 does suggest some approaches that may help mitigate these accuracy disparities such as geography-based resampling of image data sets and multi-lingual training of image-recognition models, for instance, via multi-lingual word embeddings . Such approaches may, however, still prove to be insufficient to solve the problem entirely: ultimately, the development of object-recognition models that work for everyone will likely re-quire the development of training algorithms that can learnnew visual classes from few examples and that are less susceptible to statistical variations in training data. We hope this study will help to foster research in all these directions.Solving the issues outlined in this study will allow the development of aids for the visually impaired, photo album organization software, image-search services,etc., that pro-vide the same value for users around the world, irrespective of their socio-economic status
Does Object Recognition Work for Everyone? [Terrance DeVries, Ishan Misra, Changhan Wang, Laurens van der Maaten/Facebook AI Research]
AI is worse at identifying household items from lower-income countries [James Vincent/The Verge]