Bài kiểm tra có liên kết hướng dẫn · Trung bình Cấp độ
Xu hướng kết cấu so với hình dạng trong bài kiểm tra của CNN
Thăm dò các hình ảnh xung đột tín hiệu, các quyết định về kết cấu và hình dạng, đào tạo cách điệu, các tín hiệu phụ thuộc vào nhiệm vụ và lý do tại sao độ chính xác rõ ràng là không đủ.
Câu hỏi 1 của 8
An image has a cat outline with a different animal’s texture. What can its classification reveal?
Xem trước mọi câu hỏi8 câu hỏi
Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.
- An image has a cat outline with a different animal’s texture. What can its classification reveal?
- Whether clean-set accuracy rose
- Whether the camera has metric depth
- Whether the model follows shape or texture when cues disagree
- Whether the generated image has perfect color calibration
- Why does clean-image accuracy alone not reveal shape preference?
- Shape and texture often agree on ordinary examples
- Clean examples contain no silhouettes
- Accuracy cannot be computed for CNNs
- Texture and shape always conflict on clean data
- What did the Geirhos and colleagues study observe for the CNNs it evaluated?
- Every CNN ignores global shape in all settings
- People never rely on texture
- Stylized training removes every failure mode
- They often followed texture cues in the designed conflict tests
- When might texture use be the intended behavior rather than a flaw?
- When identity should persist despite changed finish
- When classifying materials or surface defects
- When the target class is a geometric silhouette
- When texture is deliberately unrelated to the label
- A model trained with stylized images shows more shape choices. What remains necessary?
- It proves robustness to every natural shift
- It makes clean-image accuracy unnecessary
- Test on clean and target-domain photos for the actual job
- It validates every material-recognition application
- Why should style-transfer artifacts be inspected in a cue-conflict set?
- They may create a third shortcut unrelated to intended shape or texture
- Artifacts can only improve human judgments
- Artifacts are guaranteed independent of labels
- The model cannot learn generated-image artifacts
- Which measure best describes a shape-bias score?
- A direct measure of human scene understanding
- The model’s overall clean accuracy
- A guarantee against common corruptions
- The rate of shape-consistent choices on a specified cue-conflict set
- Why might maximizing shape bias harm a fabric inspector?
- Fabric defects contain no surface information
- Surface texture may contain the defects the inspector must detect
- Shape preference is always better for every task
- Texture cues never survive camera changes