Computer Vision News - February 2022

5 Agree to Disagree world just yet,” Matthew tells us. “I think it will be quite a while before we see suitable models that can be used in very sensitive scenarios, such as healthcare, where the decisions need to be perfect.” The teamused the MNIST handwritten digit dataset for baseline results before extending their experiments to more sensitive applications using MIMIC-CXR-JPG , a large publicly available chest X-ray dataset. “With the big tech companies, their applications are not very sensitive,” Bashar points out . “For them, it doesn’t matter as much if you train the model a couple of times and there are small variations in the output. If you’re classifying cats versus dogs, for example, you can afford to get it slightly wrong. But with chest X-ray, it’s very important we get it right. The same problem occurs with both models, but clearly the impact is higher for medical applications.” Noura adds : “It will take a long time to see deep learning with explainability applied in industry in the real world. One major problem is industry itself is leading most of the research, financially at least, and the easy and successful applications always get more funding. Obviously, where the work is trying to break those models and show how vulnerable they are, Bashar Awwad Shiekh Hasan Matthew Watson Noura Al Moubayed