Cynthia Dwork ICLR 2019 - Invited Talk - Highlights of Recent Developments in Algorithmic Fairness

**Concept Graph & Resume using Claude 3 Opus | Chat GPT4 | Gemini Adv | Llama 3:**

graph LR
classDef cynthia fill:#f9d4d4, font-weight:bold, font-size:14px;
classDef algorithms fill:#d4f9d4, font-weight:bold, font-size:14px;
classDef fairness fill:#d4d4f9, font-weight:bold, font-size:14px;
classDef representation fill:#f9f9d4, font-weight:bold, font-size:14px;
classDef future fill:#f9d4f9, font-weight:bold, font-size:14px;
A[Cynthia Dwork

ICLR 2019 ] --> B[Cynthia Dwork: renowned

computer scientist. 1] A --> C[Algorithms: unfair due to bias,

historical issues, features. 2] C --> D[Algorithmic unfairness: real-world

consequences. 3] A --> E[Group fairness: often

fails vs. individual. 4] E --> F[Ilvento: approximates individual

fairness metric. 5] A --> G[Multi-accuracy: group fairness

for intersectional groups. 6] A --> H[Scoring functions: unclear

probability meaning. 7] H --> I[Calibration: predicted probabilities

match frequencies. 8] G --> J[Multi-accuracy: retains set

expectations, varies without data/constraints. 9] G --> K[Capturing historical disadvantage:

consider computable sets. 10] G --> L[Multi-accuracy & calibration:

capture task-specific differences. 11] C --> M[Data: differentially expressive

for advantaged/disadvantaged. 12] A --> N[Ranking: underlies triage,

admissions, affirmative action. 13] N --> O[Fair ranking: prevent unfair

group outcomes. 14] O --> P[Multi-accuracy & calibration

prevent unfair rankings. 15] A --> Q[Focus on collected data

and measurements. 16] Q --> R[Indistinguishable examples:

assign base rate. 17] A --> S[Rich multi-calibration: predictions

as pseudo-random 'truth'. 18] A --> T[Fair representation: hide

sensitive attributes. 19] T --> U[Adversarial censoring achieves

group fairness notions. 20] T --> V[Censored representations enable

transfer learning. 21] T --> W[Censoring may identify

cross-population commonalities. 22] T --> X[Synthetic data promising for

learning common signal. 23] A --> Y[Fair algorithms alone

can't fix societal unfairness. 24] A --> Z[Metric learning breakthroughs

enable individual fairness. 25] A --> AA[Multi-calibration significant for

scoring, ranking, probabilities. 26] A --> AB[Representation and data

collection critical for fairness. 27] T --> AC[Censored representations generalize

across populations. 28] A --> AD['Superhuman' fairness remains

an open challenge. 29] A --> AE[Much work remains on

fairness and principled solutions. 30] class A,B cynthia; class C,D,M algorithms; class E,F,G,I,J,K,L,N,O,P,Q,R,S,Y,Z,AA,AD,AE fairness; class T,U,V,W,X,AB,AC representation; class H future;

ICLR 2019 ] --> B[Cynthia Dwork: renowned

computer scientist. 1] A --> C[Algorithms: unfair due to bias,

historical issues, features. 2] C --> D[Algorithmic unfairness: real-world

consequences. 3] A --> E[Group fairness: often

fails vs. individual. 4] E --> F[Ilvento: approximates individual

fairness metric. 5] A --> G[Multi-accuracy: group fairness

for intersectional groups. 6] A --> H[Scoring functions: unclear

probability meaning. 7] H --> I[Calibration: predicted probabilities

match frequencies. 8] G --> J[Multi-accuracy: retains set

expectations, varies without data/constraints. 9] G --> K[Capturing historical disadvantage:

consider computable sets. 10] G --> L[Multi-accuracy & calibration:

capture task-specific differences. 11] C --> M[Data: differentially expressive

for advantaged/disadvantaged. 12] A --> N[Ranking: underlies triage,

admissions, affirmative action. 13] N --> O[Fair ranking: prevent unfair

group outcomes. 14] O --> P[Multi-accuracy & calibration

prevent unfair rankings. 15] A --> Q[Focus on collected data

and measurements. 16] Q --> R[Indistinguishable examples:

assign base rate. 17] A --> S[Rich multi-calibration: predictions

as pseudo-random 'truth'. 18] A --> T[Fair representation: hide

sensitive attributes. 19] T --> U[Adversarial censoring achieves

group fairness notions. 20] T --> V[Censored representations enable

transfer learning. 21] T --> W[Censoring may identify

cross-population commonalities. 22] T --> X[Synthetic data promising for

learning common signal. 23] A --> Y[Fair algorithms alone

can't fix societal unfairness. 24] A --> Z[Metric learning breakthroughs

enable individual fairness. 25] A --> AA[Multi-calibration significant for

scoring, ranking, probabilities. 26] A --> AB[Representation and data

collection critical for fairness. 27] T --> AC[Censored representations generalize

across populations. 28] A --> AD['Superhuman' fairness remains

an open challenge. 29] A --> AE[Much work remains on

fairness and principled solutions. 30] class A,B cynthia; class C,D,M algorithms; class E,F,G,I,J,K,L,N,O,P,Q,R,S,Y,Z,AA,AD,AE fairness; class T,U,V,W,X,AB,AC representation; class H future;

**Resume: **

**1.-**Cynthia Dwork is a renowned computer scientist who uses theoretical computer science to address societal problems.

**2.-**Algorithms can be unfair due to biased training data, historical bias in labels, and differentially expressive features.

**3.-**Algorithmic unfairness has significant real-world consequences, such as in child protection services and recidivism prediction.

**4.-**Group fairness definitions, while popular, often fail under scrutiny compared to individual fairness.

**5.-**Ilvento's work approximates a similarity metric for individual fairness using human knowledge and learning theory.

**6.-**Multi-accuracy achieves group fairness simultaneously for intersectional groups defined by a large collection of sets.

**7.-**Scoring functions produce probabilities, but the meaning is unclear for non-repeatable events like tumor metastasis.

**8.-**Calibration in forecasting requires predicted probabilities to match observed frequencies for each predicted value.

**9.-**Multi-accuracy retains expectations for predefined sets; solutions vary without training data or additional constraints.

**10.-**Complexity theory suggests considering all efficiently computable sets to capture historically disadvantaged groups.

**11.-**Multi-accuracy and multi-calibration together aim to capture all task-specific, semantically significant differences.

**12.-**Data collected is often differentially expressive for advantaged vs. disadvantaged groups.

**13.-**Ranking underlies many applications like triage, admissions, and affirmative action strategies.

**14.-**Fair ranking should prevent obviously unfair outcomes, e.g., all of one group ranked above another.

**15.-**Multi-accuracy prevents certain unfair rankings; multi-calibration is even stronger.

**16.-**Focus should be on what data is collected and measured, as unfairness often lies there.

**17.-**Computationally indistinguishable positive and negative examples suggest assigning base rate probabilities.

**18.-**Rich multi-calibration may justify treating predictions as pseudo-random "truth" with respect to the defining sets.

**19.-**Fair representation learning aims to hide sensitive attributes while enabling standard training.

**20.-**Adversarial approaches censor representations to achieve group fairness notions like statistical parity.

**21.-**Learned censored representations can enable transfer learning to other prediction tasks.

**22.-**Censoring techniques may identify commonalities across populations for out-of-distribution generalization.

**23.-**Synthetic data experiments show promise for learning common predictive signal across populations.

**24.-**Fair algorithms alone cannot fully address societal unfairness.

**25.-**Breakthroughs in metric learning enable individual fairness.

**26.-**Multi-calibration emerged as significant for fair scoring, ranking, and understanding individual probabilities.

**27.-**Representation and data collection are critical factors in algorithmic fairness.

**28.-**Censored representations offer a path to generalizing across populations.

**29.-**Achieving truly "superhuman" fairness remains an open challenge.

**30.-**Much work remains to deeply understand fairness and develop principled, broadly applicable solutions.

Knowledge Vault built byDavid Vivancos 2024