Soften unvalidated-vs-disproven language to match the measured confound
Browse files
app.py
CHANGED
|
@@ -121,8 +121,8 @@ with gr.Blocks(title="Explainable & Auditable Alloy Defect Detection",
|
|
| 121 |
|
| 122 |
**Read the tabs in order** — they are an argument, not a menu.
|
| 123 |
The model works and explains itself; that explanation can be hijacked by a
|
| 124 |
-
meaningless artifact; and once you test the explanation directly it
|
| 125 |
-
|
| 126 |
actually hand an auditor.
|
| 127 |
"""
|
| 128 |
)
|
|
@@ -169,7 +169,8 @@ with gr.Blocks(title="Explainable & Auditable Alloy Defect Detection",
|
|
| 169 |
of **random** pixels.
|
| 170 |
|
| 171 |
A faithful explanation should collapse faster when you remove the evidence
|
| 172 |
-
it pointed at. This one did the opposite
|
|
|
|
| 173 |
Grad-CAM-ranked pixels cost only **0.546** of predicted probability, while
|
| 174 |
deleting a random 30% cost **0.718**.
|
| 175 |
"""
|
|
@@ -181,8 +182,8 @@ with gr.Blocks(title="Explainable & Auditable Alloy Defect Detection",
|
|
| 181 |
)
|
| 182 |
gr.Markdown(
|
| 183 |
"""
|
| 184 |
-
**Be precise about what
|
| 185 |
-
deletion test then
|
| 186 |
The *attribution* fell short — the *audit* worked exactly as designed,
|
| 187 |
catching something that 98% accuracy, the per-class breakdown and a
|
| 188 |
convincing-looking overlay all missed. A diagnostic that only ever confirms
|
|
|
|
| 121 |
|
| 122 |
**Read the tabs in order** — they are an argument, not a menu.
|
| 123 |
The model works and explains itself; that explanation can be hijacked by a
|
| 124 |
+
meaningless artifact; and once you test the explanation directly it cannot be
|
| 125 |
+
confirmed. What survives is a statistical guarantee, and that is what we would
|
| 126 |
actually hand an auditor.
|
| 127 |
"""
|
| 128 |
)
|
|
|
|
| 169 |
of **random** pixels.
|
| 170 |
|
| 171 |
A faithful explanation should collapse faster when you remove the evidence
|
| 172 |
+
it pointed at. This one did the opposite — which leaves it **unvalidated,
|
| 173 |
+
not disproven**: deleting the top 30% of
|
| 174 |
Grad-CAM-ranked pixels cost only **0.546** of predicted probability, while
|
| 175 |
deleting a random 30% cost **0.718**.
|
| 176 |
"""
|
|
|
|
| 182 |
)
|
| 183 |
gr.Markdown(
|
| 184 |
"""
|
| 185 |
+
**Be precise about what happened.** Grad-CAM produced the heatmap; the course's
|
| 186 |
+
deletion test then failed to confirm it for this model.
|
| 187 |
The *attribution* fell short — the *audit* worked exactly as designed,
|
| 188 |
catching something that 98% accuracy, the per-class breakdown and a
|
| 189 |
convincing-looking overlay all missed. A diagnostic that only ever confirms
|