下面是小编为大家整理的AI伦理：公平与可解释机器学习——与IBM投资者座谈会摘要,供大家参考。
Ethical
AI
-
Fair
&
Explainable Machine
Learning
DeepFin
Investor
Workshop
Summary
with
IBM
Global Quantitative & Derivatives Strategy 18
September
2020
In this research note we summarise our September 2020 DeepFin Investor
Tutorial on Fair and Explainable AI with IBM held over video conference from London. As Machine Learning and AI continue to proliferate, we explore how to remove unfair bias in AI and how to engender trust through explainable AI models. What is Machine Learning and AI? Machine Learning by its nature is a way of statistical discrimination. The discrimination becomes objectionable when it places privileged groups at a systematic advantage and certain unprivileged groups at a systematic disadvantage. Extensive evidence has shown that AI can embed human and societal bias (e.g. race, gender, caste, and religion) and deploy them at scale, consequently many algorithms are now being re-examined due to illegal bias. Trusted AI – how can we trust AI? Fundamental questions arise around how can we trust AI? And, to what level of trust can, and should we, place on AI? Forming the basis of trusted AI systems, IBM introduces four pillars: 1) Explainability: knowing and understanding
how AI models arrive at specific decisions; 2) Fairness: removing / minimising bias in the model or data; 3) Robustness: ensuring the model is safe and secure; and 4) Lineage: as models are continually evolving, we should track and maintain the provenance of data, metadata, models and test results. See cover chart. To remove unfair bias in Machine Learning, we can intervene before the models
are built (pre-processing algorithms), to a model during training (in-processing algorithms), or to the predicted labels (post-processing algorithms). Removing unfair bias in Machine Learning and explaining AI models During the practical sections of the workshop, we used Python packages from IBM in the IBM Cloud to remove unfair bias from an AI pipeline, and to help explain machine learning model predictions and data. Removing Unfair Bias in Machine Learning: using German Credit Risk data, we measured bias in the data and models, and applied a fairness algorithm to mitigate bias. By having access to the training data, we used a pre-processing algorithm to remove bias (age >
25 years): there was subsequentlynodifference in the rate of favourable outcomes received by the unprivileged group to the privileged group. Explain Machine Learning Models: how can we explain model predictions? We worked through explaining the Iris dataset predictions using SHAP, explaining German Credit Risk predictions using LIME, explain Proactive Retention decisions using TED, and analysed and explained CDC Income Data using ProtoDash – we explain each of these practical approaches in more detail within the research note. Figure 1: Four Pillars of Trusted AI Pillars of trust, woven into the lifecycle of an AI application
Big
Data
&
AI
Stategies Ayub
Hanif,
PhD
AC
(44-20)
7742-5620
ayub.hanif@jpmorgan.com
Bloomberg
JPMA
HANIF
<GO>
J.P.
Morgan
Securities
plc
Khuram
Chaudhry
AC
(44-20)
7134-6297
khuram.chaudhry@jpmorgan.com
Bloomberg
JPMA
CHAUDHRY
<GO>
J.P.
Morgan
Securities
plc
Global
Head
of
Quantitative
and Derivatives
Strategy
Marko
Kolanovic,
PhD
AC
(1-212)
622-3677
marko.kolanovic@jpmorgan.com
J.P.
Morgan
Securities
LLC
Source:
J.P.
Morgan
Quantitative
and
Derivatives
Strategy,
IBM
See page 20 for analyst certification and important disclosures, including non-US analyst disclosures. J.P. Morgan does and seeks to do business with companies covered in its research reports. As a result, investors should be aware that the firm may have a conflict of interest that could affect the objectivity of this report. Investors should consider this report as only a single factor in making their investment decision. www.jpmorganmarkets.com
Table of Contents Ethical AI – Fair & Explainable Machine Learning
.................. 3 Removing Unfair Bias in Machine Learning
............................. 6 Explore bias in the data
...............................................................................................7 Exploring bias metrics
.................................................................................................7 Select and transform features to build a model
............................................................7 Build models
...............................................................................................................8 Remove bias by reweighing data
.................................................................................8 Explain Machine Learning Models
............................................ 10 Understanding model predictions with SHAP
........................................................... 11 Understanding model predictions with LIME
........................................................... 13 Understanding model predictions with TED
............................................................. 14 Understanding data with ProtoDash
.......................................................................... 16 Takeaways
....................................................................................... 19
Ethical AI – Fair & Explainable Machine Learning
With
increased
proliferation
/ ubiquity
of
AI,
comes
increased scrutiny
How
can
we
trust
AI?
How
can
we
build
trust
in
AI?
As the hype surrounding advancements in Machine Learning and Artificial Intelligence (AI) starts to deliver, we have seen AI being increasingly used in many decisions-making applications e.g. credit, employment, admissions, sentencing and healthcare. Although Machine Learning, by its very nature, is a way of statistical discrimination, the discrimination becomes objectionable when it places certain privileged groups at a systematic advantage and certain unprivileged groups at a systematic disadvantage. Extensive evidence has shown that AI can embed human and societal bias and deploy them at scale, consequently many algorithms are now being re-examined due to illegal bias: e.g. biases in training data, due to either prejudice in labels or under- over-sampling, yields models with unwanted bias 1 . The fundamental questions thus are: how can we trust an AI system? How can an AI system explain itself? How does unfair human and societal bias leak into an AI machine? IBM describe four pillars of trust, see Figure 2, forming the basis for trusted AI systems 2 . 1. Explainability: knowing and understanding how AI models arrive at specific decisions. 2. Fairness: removing / minimising bias in the model or data. 3. Robustness: ensuring the model is safe and secure. 4. Lineage: as models are continually evolving, we should track and maintain the provenance of data, metadata, models (with hyperparameters) and test results. Figure 2: Four Pillars of Trusted AI Pillars of trust, woven into the lifecycle of an AI application
Source:
J.P.
Morgan
Quantitative
and
Derivatives
Strategy,
IBM
So, what is fairness? There are numerous definitions of fairness, many of which conflict. Consider if we were to afford positive weight to any notion of fairness through social policies may sometimes lead to reducing the well-being of every person in society 3 . Simply put, the way you define fairness impacts bias.
1
Barocas, Solon, and Andrew D. Selbst. "Big data"s disparate impact." Calif. L. Rev. 104 (2016): 671. 2
Arnold, Matthew, Rachel KE Bellamy, Michael Hind, Stephanie Houde, Sameep Mehta, A. Mojsilović, Ravi Nair et al. "FactSheets: Increasing trust in AI services through supplier"s declarations of conformity." IBM Journal of Research and Development 63, no. 4/5 (2019): 6-1. 3
Kaplow, Louis, and Steven Shavell. "The conflict between notions of fairness and the Pareto principle." American Law and Economics Review 1, no. 1 (1999): 63-77.
The following are some fairness terms used in ethical AI: Protected Attribute – an attribute that partitions a population into groups whose outcomes should have parity (e.g. race, gender, caste, and religion) Privileged Protected Attribute – a protected attribute value indicating a group that has historically been at systemic advantage Group Fairness – groups defined by protected attributes receiving similar treatments or outcomes Individual Fairness – similar individuals receiving similar treatments or outcomes Fairness Metric – a measure of unwanted bias in training data or models Favourable Label – a label whose value corresponds to an outcome that provides an advantage to the recipient In Figure 3 we show example group fairness metrics. We have two groups, unprivileged and privileged and are measuring their favourable outcome rates. In the sample scenario, 6 of the unprivileged individuals have a favourable outcome, whilst there are 7 privileged individuals with a favourable outcome. We can measure fairness from a number of different perspectives.
Statistical
Parity
Difference measures
the
difference
in positive
rates.
Disparate
Impact
expresses
the unprivileged
positive
outcomes in
relation
to
the
privileged positive
outcomes.
Equal
Opportunity
Difference measures
the
difference
in
true positives
and
false
negatives
in the
privileged
/
unprivileged groups.
Mitigate
often,
mitigate
early
Figure 3: How to measure fairness? Some group fairness metrics
Source:
J.P.
Morgan
Quantitative
and
Derivatives
Strategy,
IBM
Where can we intervene in the AI pipeline to mitigate bias? If we can modify the training data, then we use a pre-processing algorithm. If we can modify the learning algorithm, then an in-processing algorithm can be used. And if, we can only treat the learned model as a black-box and cannot modify either the training data or learning algorithm, then a post-processing algorithm can be used.
Figure 4: Where can you intervene in the AI pipeline?
Source:
J.P.
Morgan
Quantitative
and
Derivatives
Strategy,
IBM
Need
to
tradeoff
between
the bias
/
accuracy
of
the
model
with respect
to
yo...