A new way to assess AI bias in object-recognition systems

Facebook AI researchers have published the first systematic study that measures the accuracy of object-recognition systems for different communities...

24 mentions: @soumithchintala@ylecun@MIT_CSAIL@VectorInst@drsrinathsridha@hasdid@PeckRousert@AfrikanaScholar
Date: 2019/06/11 21:48

Referring Tweets

@soumithchintala Object Detection systems (commercial and academic) are trained on biased data This disproportionally affects accuracy in lower-income households and continents like Africa & Asia. Work by my colleagues at FAIR using the Dollar Street dataset from GapMinder https://t.co/gdm7Ar9f4l https://t.co/vmtkh3i4eo
@ylecun Nice new work from FAIR on assessing bias in datasets. Turns out image recognition systems will recognize everyday objects more reliably if the picture was shot in developed countries than... https://t.co/Nq2ErHND15
@MIT_CSAIL New study shows that object-detection systems are less accurate in lower-income households and on continents like Africa and Asia. https://t.co/nt672aJIHA (v/@facebookai) #ML #machinelearning #BiasInAI https://t.co/l8agwhZQhd
@VectorInst Vector PhD student Terrance DeVries @facebookai & his collaborators published a method for identifying biases in image recognition systems. Removing them will help systems more reliably recognize everyday objects in developing countries. Check it out here: https://t.co/1EO5V7kWyg
@drsrinathsridha I have always wondered how computer vision methods would work on scenes from different cultures. Objects and people look quite different, for instance, in India. This neat work from @facebookai finds that there are indeed biases and also quantifies them https://t.co/nZZBZ5qSDH
@hasdid #AINews #AI #Automated | Does object recognition work for everyone? A new method to assess bias in CV systems https://t.co/juMFpaCyN4
@oscarwijsman Does object recognition work for everyone? A new method to assess bias in CV systems https://t.co/JT3n4mr0L3 #ArtificialIntelligence https://t.co/236NPRRQ9Q
@PeckRousert New study shows that object-detection systems are less accurate in lower-income households and on continents like Africa and Asia. https://t.co/511Za36YMX #MachineLearning
@AfrikanaScholar https://t.co/aLmrhWHmyp @murageken An example of why AI field needs a lot contextualised design especially in SSA