No area of machine learning has advanced more rapidly that image classifiers in the past decade. While higher level intelligence has been harder to come by in the world of AI, image recognition feels like a baked technology, with Convolutional Neural Nets dominating the field. But a new data set from researchers at IBM and MIT called ObjectNet shows that today’s state of the art systems struggle badly when dealing with the messy real world. With objects at weird angles, or partially hidden, or covered up, even the best accuracy scores plunge from 97% to 50-55%.
- Using NLP to Bring the 7000 Languages of the World to Light
- Why Explainability is No Panacea
- If You Don’t Have Reproducibility You Don’t Have Anything
- Move Fast and Break Things? The AI Governance Dilemma
- MIT and IBM’s ObjectNet Shows Why Your Image Classifier AI Struggles at Seeing Things in the Real World