Transcript False negative

The Ethics of Image Analysis
Martin Peterson,TU/e
Outline of Lecture
1) What is ethics? And why should you care?
1) False positives vs. False negatives
2) The ROC analysis
3) Ethicals aspects
4) Ethics assignment
What is ethics
1) Descriptive ethics
2) Normative ethics
3) Metaethics
My claim: Algorithms for image analysis
contain hidden ethical assumptions.
Example: False positive vs. False negatives
The Tiger Example
In order to survive in the jungle, you build
a detector that warns for everything that
is yellow, including tigers.
- But everything edible is also yellow.
- How sensistive should the warning device be?
lightness
The Analogy Is Obvious
T1
T2
V1
V2
location
Segmentation in a noisy image. For a larger threshold value, T1 , the estimated area V1 of the ‘blood’ segment
will be lower than the estimated area V2 which is found with a lower threshold T2 . It is not a priori clear, however,
which of the two threshold values is the ‘correct’ one. The software engineer chooses a threshold without real
argument – thereby biasing the outcome of the algorithm when it is used on patient data. This may statistically
influence the change of false positive diagnostic errors in favour of false negatives, or vice versa.
False positives vs. False negatives
True positve = Device says it is
a tiger, and it is a tiger.
(You survive!)
False positve = Device says it
is a tiger, but it is not a tiger.
(You starve!)
False negative = Device says it
is not a tiger, but it is a tiger.
(You die!)
True negative = Device says it
is not a tiger, and it is not a tiger.
(You survive!)
Trade-off between FP and FN
Unfortunately, we cannot build a device that warns for all tigers
and nothing but tigers – the more sensitive the device is for
yellow things, the more bananas it will detect.
The ROC analysis 1(2)
ROC = receiver operating characteristic
• True positive rate (TPR) = TP / P = TP / (TP + FN)
• False positive rate (FPR) = FP / N = FP / (FP + TN)
• P = actual number of postives (tigers)
• N = actual number of negatives (bananas etc.)
The ROC analysis 2(2)
P = TP + FN
N = FP + TN
TP
FP
FN
TN
• True positive rate (TPR) = TP / P = TP / (TP + FN)
• False positive rate (FPR) = FP / N = FP / (FP + TN)
A Numerical Example: Detector A
FP= 28
91
FN= 37
TN= 72
109
100
100
200
TPR= 0.63
FPR=0.28
63 / 63+37
28 / 28+72
• TP=
63
A Numerical Example: Detector B
FP=77
154
FN=23
TN=23
46
100
100
200
TPR=0.77
FPR=0.77
77 / 77 + 23
77 / 77 + 23
• TP=77
The area under the ROC curve is
sometimes used as a score
The radar on my sailing boat….
Ethical Aspects of the ROC analysis
Algorithms,
(i) sometimes comprise ethical assumptions, and
(ii) influence our decisions,
(iii) but algorithms are not moral agents.
Why Do Alogorithms Comprise Ethical
Assumptions?
VALUE-JUDGEMENT:
A proposition is a value-judgement if and only if it expresses a view
on how things ought to be, or what is good or bad, or desirable or
undesirable.
Do all algorithms comprise value-judgments? No. But some do.
DEFINITION
•
An algorithm comprises a value-judgment if (but possibly not only
if) two or more persons who accept different value-judgments would
have a rational reason to design or use the model differently.
The limits of using area under the ROC
curve an ethical criterion
The Goal of Science vs.
The Goal of Medicine
Science:
Seek the truth. Build upon accumulated knowledge. Be conservative
– a false negative is better than a false positive?
Medicine:
Protect the patient. Better to be safe than sorry – a false positive is
better than a false negative?
•
.
A Moral Principle: Transparency
Transparency
Another Moral Principle:
The Precautionary Principle
”Where there are threats of serious or irreversible
damage, lack of full scientific certainty shall not be
used as a reason for postponing cost-effective
measures to prevent environmental degradation.”
(UNCED)
“When an activity raises threats of harm to the
environment or human health, precautionary
measures should be taken even if some cause and
effect relationships are not fully established
scientifically.” (Wingspread Conference)
What kind of principle is PP?
• A legal principle
• A normative principle
• An epistemic principle
.
•
Normative interpretations of PP: What should we
do?
•
Epistemic interpretations of PP: What should we
believe?
- Reversed burden of proof?
- If experts disagree, trust the pessimistic ones?
- False positive/negative errors
An objection to the epistemic version of PP:
•
•
•
•
You live in the jungle.
Tigers are yellow and black.
Everything edible in the jungle is also yellow (Bananas)
To protect yourself against tigers you build a device that detects and warns for
everything that is yellow.
• The detector is very sensitive  Many false positives
--------------------------------------------------------• Good news: Because of the detector you will not be killed by
a tiger.
• Bad news: You will starve to death, because you will never
find anything to eat.
----------------------------------------------------------• Conclusion: It is far from clear that it is in general better to
prefer false positives over false negatives.
Assignment
Include 2 A4 pages in the final report, in which you:
1) Explain what an ROC analysis is and why it is important to
your project.
2) Discuss the ethical aspects of the choice between a high
TPR and a low FPR. What would a reasonable choice be for
your project?
3) Motivate why the TPR and FPR of your software is at it is.