Beyond Accuracy: What Happens When AI Meets Real-World Data? An AI model can perform exceptionally well on its test dataset and still struggle with images from a different source. Why does this happen, and how can we evaluate models more meaningfully? In this session, Devanandha Vellaramkuzhiyil Shaji shares lessons from an independent wheat rust detection project, exploring transfer learning, cross-dataset testing and explainable AI using Grad-CAM. The talk examines how training-data composition affects generalisation, what visual explanations can—and cannot—tell us, and why reporting limi…