45 Making MetaDataCount Computer Vision News Computer Vision News Since the same patients are tracked over an extended period of time, it is essential to prevent data leakage. If the data is not split carefully, patients in the evaluation data may also appear in the training set, resulting in artificially high performance scores. The results she presented show recordwise split achieved the highest accuracy, but evaluation on a hold-out test set revealed a large drop in accuracy. GradCAM visualization techniques were used to get a better insight on why the model was underperforming. These visualizations showed the model was learning shortcuts unrelated to the task at hand. Her entire talk is in the video here above. Would you be interested in joining our next webinar to improve your own knowledge or have the opportunity to asks experts questions? Please look at our website and sign-up to the newsletter!
RkJQdWJsaXNoZXIy NTc3NzU=