Minor bugs in 13.md

slides/13/13.md

@@ -37,22 +37,22 @@ which seems morally good or neutral…
 
 # Why ML Ethics? (2)
 
-* There are high-stake applications of ML: decision making (in public and business administration)
+* There are high-stake applications of ML: decision making (in public and business administration).
 
 <br />
 
 ~~~
-* People tend to consider algorithm outputs as objective
+* People tend to consider algorithm outputs as objective.
 
 <br />
 
 ~~~
-* There are bad actors (military use, massive surveillance, disinformation, ...)
+* There are bad actors (military use, massive surveillance, disinformation, ...).
 
 <br />
 
 ~~~
-* Inherent problems of ML: Good intentions might lead to unintended harm (biases, lack of explainability)
+* Inherent problems of ML: Good intentions might lead to unintended harm (biases, lack of explainability).
 
 
 ---
@@ -63,36 +63,36 @@ class: section
 ---
 # What is Ethics
 
-Ethics or moral philosophy is a branch of philosophy that "involves
+Ethics or moral philosophy is a branch of philosophy that “involves
 systematizing, defending, and recommending concepts of right and wrong
-behavior".
+behavior”.
 
 <br />
 
 ~~~
-* Simply taken: Study of what is **right and wrong**
+* Simply taken: Study of what is **right and wrong**.
 
 ~~~
 * Several major theoretical frameworks:
 
-  * **Deontological** ethics – rule/principle-driven, good action follows rules
+  * **Deontological** ethics – rule/principle-driven, good action follows rules.
 
-  * **Consequentialist** ethics / utilitarianism  – consequences matter, good action has good consequences
+  * **Consequentialist** ethics / utilitarianism  – consequences matter, good action has good consequences.
 
-  * **Virtue** ethics (encourage curiosity, creativity, solidarity, etc.) – good action is what a virtuous person would do
+  * **Virtue** ethics (encourage curiosity, creativity, solidarity, etc.) – good action is what a virtuous person would do.
 
-  * **Contract** ethics (everything is a social contract) – good action is what people implicitly agreed upon
+  * **Contract** ethics (everything is a social contract) – good action is what people implicitly agreed upon.
 
-  * **Care ethics** – relationships and responsibilities matter, good action nurtures care and empathy for others
+  * **Care ethics** – relationships and responsibilities matter, good action nurtures care and empathy for others.
 
 ---
 # Deontological Ethics
 
-* Focuses on the **inherent nature of actions** rather than their consequences
+* Focuses on the **inherent nature of actions** rather than their consequences.
 
-* Involves adhering to predefined **rules and principles** (e.g., the Universal Declaration of Human Rights, the Ten Commandments, Kant's Categorical Imperative)
+* Involves adhering to predefined **rules and principles** (e.g., the Universal Declaration of Human Rights, the Ten Commandments, Kant's Categorical Imperative).
 
-* In the ML context, it typically means principles like: beneficence, non-malevolence, privacy, non-discrimination, autonomy + informed consent
+* In the ML context, it typically means principles like: beneficence, non-malevolence, privacy, non-discrimination, autonomy + informed consent.
 
 <br />
 ~~~
@@ -125,9 +125,9 @@ behavior".
 ---
 # Utilitarian Ethics
 
-* An ethical theory that emphasizes the maximization of **overall happiness or well-being** or minimizing harm
+* An ethical theory that emphasizes the maximization of **overall happiness or well-being** or minimizing harm.
 
-* It focuses on **consequences of actions**
+* It focuses on **consequences of actions**.
 
 * Good decisions = decisions that lead to the greatest overall positive impact.
 
@@ -154,9 +154,9 @@ behavior".
 ~~~
 <br />
 
-* Implicitly behind most work on ML ethics that focuses on harmful consequences for various social groups
+* Implicitly behind most work on ML ethics that focuses on harmful consequences for various social groups.
 
-* Gets tricky once we consider very low-probability events with very high impact (special case longtermism)
+* Gets tricky once we consider very low-probability events with very high impact (special case longtermism).
 
 ---
 # Different Theories, Different Ideas
@@ -188,17 +188,17 @@ class: section
 ---
 # Stages of ML development
 
-Ethical problems might emerge in all stages of ML system development
+Ethical problems might emerge in all stages of ML system development.
 
-* **Problem definition** – some tasks are inherently problematic
+* **Problem definition** – some tasks are inherently problematic.
 
-* **Data collection** – biases in data, unethical collection
+* **Data collection** – biases in data, unethical collection.
 
-* **Model development** – design choices (i.e., most of this course)
+* **Model development** – design choices (i.e., most of this course).
 
-* **Model evaluation** – metrics do not cover important things
+* **Model evaluation** – metrics do not cover important things.
 
-* **Model deployment** – use outside of original scope, feedback loops
+* **Model deployment** – use outside of original scope, feedback loops.
 
 ---
 # Problem Definition
@@ -218,12 +218,12 @@ Many schools in the US use automatic essay scoring for Graduate Record Examinati
 ![w=30%,f=right](essays.png)
 ![w=30%,f=right](essay_map.png)
 
-* **Lack of transparency**: students have the right to know why they were accepted
+* **Lack of transparency**: students have the right to know why they were accepted.
 
-* Allows **metric gaming** if you guess what the features might be <small>(so even the utility is low)</small>
+* Allows **metric gaming** if you guess what the features might be <small>(so even the utility is low)</small>.
 
 ---
-# Recidivism prediction: COMPAS
+# Recidivism Prediction: COMPAS
 
 ![w=30%,f=right](compas.png)
 
@@ -242,31 +242,31 @@ Many schools in the US use automatic essay scoring for Graduate Record Examinati
 ~~~
 
 **Deontology:** violates the right to a fair trial, equality before the law, lack of transparency
-⇒ *Morally bad*
+⇒ *Morally bad*.
 
 ~~~
 
 **Utilitarianism:** the benefits <small>(the state saves money that can be use elsewhere)</small> are smaller than the harm <small>(lack of justice)</small>
-⇒ *Morally bad* <br /> <small>(But what would be the suitable metric to compare money and fairness?)</small>
+⇒ *Morally bad*. <br /> <small>(But what would be the suitable metric to compare money and fairness?)</small>
 
 ---
 # Data Collection & Biases in Data
 
 ![w=30%,f=right](gendershades.png)
 
 * Representation bias: The data might not be representative of the population
-  (missing minorities, poor people, ...)
+  (missing minorities, poor people, ...).
 
 ~~~
 * Data (and especially text) from the Internet does not represent the world as
-  it is (only those who have access and are loud) and the world as it should be
+  it is (only those who have access and are loud) and the world as it should be.
 
 ![w=30%,f=right](imagenet.png)
 
 ~~~
-* Historical bias: Inequalities from the past when the data was created are preserved in the datasets
+* Historical bias: Inequalities from the past when the data was created are preserved in the datasets.
 
-* Copyright issues, especially with generative models
+* Copyright issues, especially with generative models.
 
 ![w=100%,h=center](female_doctor.png)
 
@@ -275,54 +275,54 @@ Many schools in the US use automatic essay scoring for Graduate Record Examinati
 
 ### Crowdsourcing
 
-* People are hired to do the job the ML model will do
+* People are hired to do the job the ML model will do.
 
-* Not well paid (often in third world countries), monotonous work, occasionally causing psychological harm
+* Not well paid (often in third world countries), monotonous work, occasionally causing psychological harm.
 
-* Gig economy: what was originally meant as earning extra cash becomes a full-time job without labor protection
+* Gig economy: what was originally meant as earning extra cash becomes a full-time job without labor protection.
 
 <small>Crawford, Kate. The atlas of AI: Power, politics, and the planetary costs of artificial intelligence. Yale University Press, 2021. Chapter 2.</small>
 
 ~~~
 ### Log mining and user data collection
 
-* Training data is collected from users that have no other choice than to provide data by using services <small>(not using them and keeping social/work/political life at the same is impossible)</small>
+* Training data is collected from users that have no other choice than to provide data by using services <small>(not using them and keeping social/work/political life at the same is impossible)</small>.
 
-* Nontransparent transaction: user gets service (for free or paid) and provides data
+* Nontransparent transaction: user gets service (for free or paid) and provides data.
 
 <small>Couldry, Nick, and Ulises A. Mejias. The costs of connection: How data is colonizing human life and appropriating it for capitalism. Stanford University Press, 2020.</small>
 
 ---
 # Model Development
 
-* Discretization of outputs might lead to bias amplification <br />
+* Discretization of outputs might lead to bias amplification. <br />
   <small>Men Also Like Shopping: Reducing Gender Bias Amplification using Corpus-level Constraints (Zhao et al., EMNLP 2017)</small>
 
 ~~~
 
-* Larger models are more prone to overfitting: might lead to memorization of very specific patterns (privacy issues, remembering particular names) <br />
+* Larger models are more prone to overfitting: might lead to memorization of very specific patterns (privacy issues, remembering particular names). <br />
 <small>Are Large Pre-Trained Language Models Leaking Your Personal Information? (Huang et al., Findings EMNLP 2022)</small>
 
 ~~~
 
-* Distilled models are prone to stereotyping <br />
+* Distilled models are prone to stereotyping. <br />
 <small>Why Knowledge Distillation Amplifies Gender Bias and How to Mitigate from the Perspective of DistilBERT (Ahn et al., GeBNLP 2022)</small>
 
 ~~~
 
-* Models might learn protected attributes by proxies (e.g., ethnicity from name, school) <br /> <small>Adversarial Removal of Demographic Attributes Revisited (Barrett et al., EMNLP-IJCNLP 2019)</small>
+* Models might learn protected attributes by proxies (e.g., ethnicity from name, school). <br /> <small>Adversarial Removal of Demographic Attributes Revisited (Barrett et al., EMNLP-IJCNLP 2019)</small>
 
 ---
 # Model Evaluation
 
-* Metrics might not capture everything we need
+* Metrics might not capture everything we need.
 
-  * E.g., translation fluency does not capture gender bias
+  * E.g., translation fluency does not capture gender bias.
 
-  * Micro-averaging might hide bad performance for specific user groups (typically minorities)
+  * Micro-averaging might hide bad performance for specific user groups (typically minorities).
 
 ~~~
-* Human Resources: employment recommendation based on CV
+* Human Resources: employment recommendation based on CV.
 
   * Precision: The business implies optimizing for precision – you only
     recommend few candidates and they need to be the good ones.
@@ -331,7 +331,7 @@ Many schools in the US use automatic essay scoring for Graduate Record Examinati
     against gender, age, ethnicity, etc.
 
 ---
-# Proxy metrics optimizing something else
+# Proxy Metrics Optimizing Something Else
 
 ![w=30%,f=right](hal9000.jpg)
 
@@ -347,21 +347,21 @@ Many schools in the US use automatic essay scoring for Graduate Record Examinati
 
 ~~~
 * Platforms like YouTube use watch time as a proxy for content quality (and
-  btw. more watch time brings them more money)
+  btw. more watch time brings them more money).
 
 * Non-profit [algotransparency.org](https://www.algotransparency.org) monitors
   stats on YouTube recommendations: 2016-2018 most recommended videos
   supporting alternative narrative on political events (US and French
-  elections, mass shootings)
+  elections, mass shootings).
 
-* Presumably, this was the type of content maximizing the watch time
+* Presumably, this was the type of content maximizing the watch time.
 
 <br />
 
 <small>https://guillaumechaslot.medium.com/how-algorithms-can-learn-to-discredit-the-media-d1360157c4fa</small>
 
 ---
-# Use of model in practice & Feedback loops
+# Use of Model in Practice & Feedback Loops
 
 ### Mismatch of train/test data and use in practice
 
@@ -399,68 +399,68 @@ class: section
 # Review of the Semester
 
 ---
-# Theoretical concepts
+# Theoretical Concepts
 
 ### Basic statistic
 
-Bernoulli distribution, Categorical distribution, Normal
+Bernoulli distribution, categorical distribution, Normal
 Distribution, descriptive statistics (mean, variance, correlation), Maximum
-Likelihood Estimation, Bayes Theorem
+Likelihood Estimation, Bayes' Theorem.
 
 ### Information theory basics
 
-Entropy, Conditional Entropy, Cross-Entropy, KL-Divergence, Mutual Information
+Entropy, Conditional Entropy, Cross-Entropy, KL-Divergence, Mutual Information.
 
-* Training = minimize how suprised we are from the data
-* Maximum entropy principle = a view on generalization: do not bring in additional assumptions
+* Training = minimize how suprised we are from the data.
+* Maximum entropy principle = a view on generalization: do not bring in additional assumptions.
 
 ### Optimization
 
 Set derivative to zero, Lagrange multipliers for additional constraints, numerical optimization with SGD
-and second-order methods
+and second-order methods.
 
 ---
-# Machine learning methodology
+# Machine Learning Methodology
 
 ### Working with data
 
-* **Data annotation**: inter-annotator agreement (correlation, Cohen's alpha)
+* **Data annotation**: inter-annotator agreement (correlation, Cohen's kappa).
 
-* **Features**: numerical/categorical features, polynomial features, TF-IDF, use pre-trained embeddings, representation learning
+* **Features**: numerical/categorical features, polynomial features, TF-IDF, use pre-trained embeddings, representation learning.
 
-* **Normalization**: min-max scaling, standardization, whitening
+* **Normalization**: min-max scaling, standardization, whitening.
 
 ### Training and Evaluation
 
-* **Data splits**: optimization for unseen data, train, validation, test split
+* **Data splits**: optimization for unseen data, train, validation, test split.
 
-* **Overfitting and regularization**: early stopping + reading learning curves, $L^2$-regularization, dropout in MLP, prior in Bayesian models
+* **Overfitting and regularization**: early stopping + reading learning curves, $L^2$-regularization, dropout in MLP, prior in Bayesian models.
 
-* **Evaluation metrics**: accuracy, mean squared error, precision/recall/F-score, correlation, hypotheses testing
+* **Evaluation metrics**: accuracy, mean squared error, precision/recall/F-score, correlation, hypotheses testing.
 
 ---
-# Machine learning models
+# Machine Learning Models
 
 ### Geometric intuition
 
-Linear regression, Perceptron, Nearest Neighbors Classification and Regression, SVD, $k$-Means clustering
+Linear regression, Perceptron, Nearest Neighbors Classification and Regression, SVD, $k$-Means clustering.
 
 ### Probabilistic intuition
 
-Linear regression, logistic regression, Multi-layer Perceptron, Naive Bayes, PCA
+Linear regression, logistic regression, Multilayer Perceptron, Naive Bayes, PCA.
 
 ### Decision trees
 
-Random forest, Gradient boosted decision trees
+Random forest, gradient-boosted decision trees.
 
 ---
-# Course Objectives: What you hopefully learned
+# Course Objectives: What You Hopefully Learned
 
 After this course you should…
 
 - Be able to reason about tasks/problems **suitable for ML**
    - Know when to use classification, regression and clustering
-   - Be able to choose from this method Linear and Logistic Regression,
+   - Be able to choose from these methods: Linear and Logistic Regression,
      Multilayer Perceptron, Nearest Neighbors, Naive Bayes, Gradient Boosted Decision
      Trees, $k$-means clustering