Case context: A product team is building an automated image-tagging tool for a large photo archive. Their current design first converts each image into a fixed set of handcrafted texture and color descriptors, then passes those descriptors into a second model that predicts a short list of labels, which are finally expanded into full tags. Even after training on much more labeled data, the system stops improving and remains far below expert performance.

Question: Based on the system design, what is the main architectural reason for the performance ceiling, and how do the chosen components create that limit?

Sample answer: The main problem is that the system depends on hand-designed intermediate representations, which limit what the model can learn end to end. The handcrafted texture and color descriptors compress the image and can throw away details that matter for tagging. The intermediate label prediction stage is also an artificial representation that may not match the true structure of the tagging task. Because both steps are fixed human-designed bottlenecks, they cap the best performance the system can reach even if more data is added.

Key points:

• Hand-engineered intermediate components can cap maximum performance.
• Handcrafted descriptors may discard useful information from the original image.
• An intermediate label stage can be an imperfect representation of the task.
• These forced representations create a bottleneck that extra data cannot remove.

Rubric: The answer should identify that hand-engineered components are limiting performance by discarding information and forcing an imperfect intermediate representation.