CMSC 304

Social and Ethical Issues in Information Technology

Evaluating AI

AI & Machine Learning

Model Cards

Quiz Review

Quiz Review

Announcements

• Check your email for the one I sent about your peer feedbacks and due date for revisions!
• Paper grades will be replaced once revisions are in
  • Check with TA if you have a question that's not answered in the rubric or comments
• This week we return back to our usual rhythm. As a reminder our usual policies are:​
  • Pls don't bring laptops and do other work in class
  • Short reading quiz on Monday (today!)
  • (short) reflection journal due this Friday
  • The reading for next Monday is lorge, I recommend starting early
    • I've left some notes in the margins for items to focus on, and also some general questions in the bookmark

AI vs. ML vs. DL

https://blogs.nvidia.com/blog/whats-difference-artificial-intelligence-machine-learning-deep-learning-ai/

• AI has been a topic of imagination and research since 1956, where computer scientists formally established the field.

• Early machine learning included artificial neural networks inspired by biology of our brains: interconnections between neurons. Basically laughed at until 2012.

• Computer vision emerged as a major application area for machine learning, though it initially required significant hand-coding.

• Deep learning starting in 2015 made advances on all of the above. This recent AI explosion driven by the availability of GPUs.

Algorithmic Decision-making

• What is an algorithm?
  • A set of rules and procedures that leads to a decision
  • Could be machine learning, or simply control flow logic (rule-based if-then-else)
• What can algorithms do for us?
  • They improve accuracy and efficiency over human decision-making
  • Algorithms can make decisions faster and more consistently than humans
  • We can just create the model to be optimal, since it's just numbers (right?)
• How can algorithms harm us?
  • "An algorithm is only as good as ..."
    • its training data
    • its designer
  • Widespread rapid adoption can encode bias, threaten fairness, erase privacy, transparency, and due process

Burton, Emanuelle; Goldsmith, Judy; Mattei, Nicholas; Siler, Cory; Swiatek, Sara-Jo. Computing and Technology Ethics: Engaging through Science Fiction (pp. 117-118). MIT Press.

The promise of "algorithms"

• Many modern systems rely on some combination of data, algorithms, AI/ML to make decisions and manage complex processes
  • sometimes you'll see this called buzzwords "data-driven" or "AI-powered"
• Automated decision-making systems are typically meant to improve and enhance human decision-making
  • medical diagnosis
  • bail and sentencing
  • finance and lending
  • product recommendation
• The implication is that human decision-making is flawed, and automated systems can remove human bias, inefficiency, limitation for information processing, speed, etc.

Fairness in Algorithmic Decision-making

• Algorithms require formalization of what to compute
  • you need to actually write an equation of your goal i.e. an objective function​
    • What is the cost of making each decision?

• You also need to write an equation for evaluating the results for accuracy and fairness
  • Is accuracy the same thing as fairness?

https://web.stanford.edu/class/cs182/

Criteria for Algorithmic Accountability

1. Does it work? (Efficiency to achieve public safety)
2. Can people understand how it works? (Transparency)
3. Is it fair? (Fairness)
4. Can people appeal its judgment? (Due Process)
5. Does it use information that an individual might reasonably expect to remain private? (Privacy)

https://web.stanford.edu/class/cs182/

What does Artificial Intelligence have to do with ethics?

Today, many (most?) ethical decisions are written in software

• ranking
• optimization​
• recommendations
• content moderation
• When you start dealing with multiple, often competing, objectives or try to account for intangibles like “freedom” and “well-being,” a satisfactory mathematical solution doesn’t always exist.

https://web.stanford.edu/class/cs182/

Design and Values

Catchphrases:

• Whenever we create a computer system, we select options from a set of design choices
• Our design choices encode a set of values
  • e.g. I could make this colorblind friendly, but men have anough advantages already
  • e.g. I could 
• sometimes your design choices will be limited by regulation
  • e.g. dark patterns GDPR
  • unsubscribe
• whose values do we use, and why?
• Do Artifacts Have Politics?
• Data is Not Neutral
• some say that exploring and discussing these values is part of your civic duty as a computing professional

Design and Values

• To move from intuition to ethic, we need to surface our values
• In order to examine and evaluate competing values, we have to make them explicit and make them strange
• how do different design choices change outcomes?
• When you start dealing with multiple, often competing, objectives or try to account for intangibles like “freedom” and “well-being,” a satisfactory mathematical solution doesn’t always exist.
• remember:
  • goal is NOT to calculate the right answer
  • goal is NOT to persuade until unanimous
  • goal is to systematically surface and weigh benefits and harms
    • ​ultimately make a decision you can live with

What is Machine Learning?

• We have a wide range of students in this class!
  • Sophomores, Juniors, Seniors
  • Some have taken courses in Machine Learning, AI, Data Science, some haven't
  • Some people have had internships where they work on these areas, some haven't
  • Some people have done some self learning / personal projects in these areas
• I want to make sure this module can be understood by everyone, regardless of your pre-reqs, work history, etc.
  • SO: if you're already a pro, you may already have experience with these topics
    • As you're following along with each topic, think about something that you struggled with when you first learned this stuff
    • Think about how you could help a classmate that is feeling a bit lost
    • If you learn something new, jot it down!
• There will be a quiz question next week on this stuff (there's a reading if you don't catch it all today) 

What is Machine Learning?

• The process is the same as "human learning"
• We just need to re-align our vocab

What is Machine Learning?

• To mathify it, let's use an analogy: fitting a line to some points

What is Machine Learning?

• To mathify it, let's use an analogy: fitting a line to some points

What is Machine Learning?

• To mathify it, let's use an analogy: fitting a line to some points

What is Machine Learning?

• To mathify it, let's use an analogy: fitting a line to some points

What is Machine Learning?

• To mathify it, let's use an analogy: fitting a line to some points

question: is this a "good" model?

How could you evaluate this model to find out?

What is Machine Learning?

• To mathify it, let's use an analogy: fitting a line to some points

question: is this a "good" model?

What is Machine Learning?

• To mathify it, let's use an analogy: fitting a line to some points

question: is this a "good" model?

more interesting measures:

on average, how much error can we expect in our prediction?

how much variation can we expect in that error?

What is Machine Learning?

• One of the tasks in machine learning is regression
  • ​in statistics and machine learning, regression means predicting a quantitative dependent variable y based on one or more independent variables x (also called features).
  • examples: house price prediction, traffic flow prediction
• Another common task is classification
  • classification means predicting a categorical label or class L for a given input based on its features x, y...
  • examples:  Classifying emails as "spam" or "not spam", identifying objects in images (e.g., "cat," "dog," "car"), assigning multiple topics or tags to a document

image from https://paperswithcode.com/task/classification

What is Machine Learning?

What is Machine Learning?

cat zone

dog zone

What is Machine Learning?

cat zone

dog zone

What is Machine Learning?

cat zone

dog zone

question: are these "good" models?

How could you evaluate these models to find out which is better?

How Can We Evaluate Algorithms?

• Let's look at our data again and try a few decision boundaries

How Can We Evaluate Algorithms?

• Let's look at our data again and try a few decision boundaries

instead of measuring error, we can look at accuracy

How Can We Evaluate Algorithms?

• Let's look at our data again and try a few decision boundaries

instead of measuring error, we can look at accuracy

How Can We Evaluate Algorithms?

• Let's look at our data again and try a few decision boundaries

instead of measuring error, we can look at accuracy

How Can We Evaluate Algorithms?

• Let's look at our data again and try a few decision boundaries

instead of measuring error, we can look at accuracy

How Can We Evaluate Algorithms?

• Let's look at our data again and try a few decision boundaries

instead of measuring error, we can look at accuracy

How Can We Evaluate Algorithms?

• Let's look at the model's confusion between classes

instead of measuring error, we can look at accuracy

we can also ask, which classes get confused with others?

How Can We Evaluate Algorithms?

• Let's look at the model's confusion between classes

How Can We Evaluate Algorithms?

• Let's look at the model's confusion between classes

How Can We Evaluate Algorithms?

• Let's look at the model's confusion between classes

How Can We Evaluate Algorithms?

• Let's look at the model's confusion between classes

How Can We Evaluate Algorithms?

• Let's look at the model's confusion between classes

How Can We Evaluate Algorithms?

• Let's look at the model's confusion between classes

• False Negative = error of omission. Item was predicted as "dog," was in fact a "cat"
• False Positive = error of commission. Item was predicted as "cat," was in fact a "dog"

Now you try!

Step 1: fill in the data table

Step 1: fill in the data table

Step 2: complete the confusion matrix

Step 2: complete the confusion matrix

Step 1: fill in the data table

Step 2: complete the confusion matrix

Step 1: fill in the data table

7

Step 2: complete the confusion matrix

Step 1: fill in the data table

FN = 1

7

Step 2: complete the confusion matrix

Step 1: fill in the data table

FN = 1

FN = 1

7

Step 2: complete the confusion matrix

7

7

Step 1: fill in the data table

FN = 1

FN = 1

FN = 5

Step 2: complete the confusion matrix

7

7

Step 1: fill in the data table

FN = 1

FN = 1

FN = 5

FP = 5

Step 2: complete the confusion matrix

7

7

Step 1: fill in the data table

FN = 1

FN = 1

FN = 5

FP = 5

FP = 1

If you're not a fan of this whole summing across the rows and columns thing...

instead you can break them down into 2x2's and then sum them in the 3rd dimension

If you're not a fan of this whole summing across the rows and columns thing...

instead you can break them down into 2x2's and then sum them in the 3rd dimension

FN = 1

FN = 1

FN = 5

FP = 5

FP = 1

FP = 1

If you're not a fan of this whole summing across the rows and columns thing...

instead you can break them down into 2x2's and then sum them in the 3rd dimension

FN = 1

FN = 1

FN = 5

FP = 5

FP = 1

FP = 1

If you're not a fan of this whole summing across the rows and columns thing...

instead you can break them down into 2x2's and then sum them in the 3rd dimension

FN = 1

FN = 1

FN = 5

FP = 5

FP = 1

FP = 1

Interpreting Evaluation Results

predicted labels

Interpreting Evaluation Results

To interpret the results, we need to know "out of how many?" for each class

• For False Positives, we want to know how much each class was over-reported, out of how many opportunities?
• For False Negatives, we want to know how many we missed, out of how many opportunities?
• For True Positives, we want to know how many we got right, out of all opportunities?

predicted labels

Summary of the possible outcomes

• True Positives (TP): The number of positive instances correctly classified as positive. E.g., predicting "shoe" when it actually is a shoe.
• False Positives (FP): The number of negative instances incorrectly classified as positive. E.g., predicting "shoe" when it actually is not a shoe.
• True Negatives (TN): The number of negative instances correctly classified as negative. E.g., predicting it is not a shoe when it actually is not a shoe.
• False Negatives (FN): The number of positive instances incorrectly classified as negative. E.g., predicting as "tree" or "bicycle" when it actually is a shoe.

Summary of the possible outcomes

When we have class imbalance, we should report these as percentages or rates, or "out of how many?"

For reasons, we call the True Positive Rate Recall

• True Positives (TP): The number of positive instances correctly classified as positive. E.g., predicting "shoe" when it actually is a shoe.
• False Positives (FP): The number of negative instances incorrectly classified as positive. E.g., predicting "shoe" when it actually is not a shoe.
• True Negatives (TN): The number of negative instances correctly classified as negative. E.g., predicting it is not a shoe when it actually is not a shoe.
• False Negatives (FN): The number of positive instances incorrectly classified as negative. E.g., predicting as "tree" or "bicycle" when it actually is a shoe.

Summary of the possible outcomes

When we have class imbalance, we should report these as percentages or rates, or "out of how many?"

For reasons, we call the True Positive Rate Recall

• True Positives (TP): The number of positive instances correctly classified as positive. E.g., predicting "shoe" when it actually is a shoe.
• False Positives (FP): The number of negative instances incorrectly classified as positive. E.g., predicting "shoe" when it actually is not a shoe.
• True Negatives (TN): The number of negative instances correctly classified as negative. E.g., predicting it is not a shoe when it actually is not a shoe.
• False Negatives (FN): The number of positive instances incorrectly classified as negative. E.g., predicting as "tree" or "bicycle" when it actually is a shoe.

Recall uses all positives as denominator

Summary of the possible outcomes

When we have class imbalance, we should report these as percentages or rates, or "out of how many?"

For reasons, we call the True Positive Rate Recall

• True Positives (TP): The number of positive instances correctly classified as positive. E.g., predicting "shoe" when it actually is a shoe.
• False Positives (FP): The number of negative instances incorrectly classified as positive. E.g., predicting "shoe" when it actually is not a shoe.
• True Negatives (TN): The number of negative instances correctly classified as negative. E.g., predicting it is not a shoe when it actually is not a shoe.
• False Negatives (FN): The number of positive instances incorrectly classified as negative. E.g., predicting as "tree" or "bicycle" when it actually is a shoe.

Recall uses all positives as denominator

Precision uses predicted positives as denominator

Summary of the possible outcomes

When we have class imbalance, we should report these as percentages or rates, or "out of how many?"

For reasons, we call the True Positive Rate Recall

• True Positives (TP): The number of positive instances correctly classified as positive. E.g., predicting "shoe" when it actually is a shoe.
• False Positives (FP): The number of negative instances incorrectly classified as positive. E.g., predicting "shoe" when it actually is not a shoe.
• True Negatives (TN): The number of negative instances correctly classified as negative. E.g., predicting it is not a shoe when it actually is not a shoe.
• False Negatives (FN): The number of positive instances incorrectly classified as negative. E.g., predicting as "tree" or "bicycle" when it actually is a shoe.

Recall uses all positives as denominator

Precision uses predicted positives as denominator

Summary of the possible outcomes

When we have class imbalance, we should report these as percentages or rates, or "out of how many?"

For reasons, we call the True Positive Rate Recall

Recall uses all positives as denominator

Precision uses predicted positives as denominator

Your turn! Calculate the Precision & Recall

When we have class imbalance, we should report these as percentages or rates, or "out of how many?"

Recall uses all positives as denominator

Precision uses predicted positives as denominator

Your turn! Calculate the Precision & Recall

When we have class imbalance, we should report these as percentages or rates, or "out of how many?"

Recall uses all positives as denominator

Precision uses predicted positives as denominator

Your turn! Calculate the Precision & Recall

When we have class imbalance, we should report these as percentages or rates, or "out of how many?"

Recall uses all positives as denominator

Precision uses predicted positives as denominator

Your turn! Calculate the Precision & Recall

When we have class imbalance, we should report these as percentages or rates, or "out of how many?"

Recall uses all positives as denominator

Precision uses predicted positives as denominator

A Metric Ton of Metrics

https://en.wikipedia.org/wiki/Precision_and_recall

Documenting Classifier Performance and Trade-offs

• Precision Recall Curves
• Model Cards

Visualizing Classifier Trade-offs

https://scikit-learn.org/1.5/auto_examples/model_selection/plot_precision_recall.html

• We computed precision and recall for just one instance of this model
• In real world settings, there are often multiple parameters that we can tune to optimize the number of False Positives, False Negatives, and True Positives that we get
  • stuff in the actual model's equation are called hyperparameters
  • most of the time, we get a confidence value as output, which needs to be turned into a categorical label L
    • probability of an observation belonging to each class determines which label it gets

Visualizing Classifier Trade-offs

• We computed precision and recall for just one instance of this model
• In real world settings, there are often multiple parameters that we can tune to optimize the number of False Positives, False Negatives, and True Positives that we get
  • stuff in the actual model's equation are called hyperparameters
  • most of the time, we get a confidence value as output, which needs to be turned into a categorical label L
    • probability of an observation belonging to each class determines which label it gets
    • we can "sweep a curve" over all possible thresholds, and plot the resulting P and R

https://scikit-learn.org/1.5/auto_examples/model_selection/plot_precision_recall.html

Course Rhythm

Text

• bullet
• bullet

Values to sweep?

PR Curves Example

https://stripe.com/en-gi/guides/primer-on-machine-learning-for-fraud-protection

Model Cards

• Model cards are used by algorithm developers and companies to summarize the key performance metrics of their developed system
  • essential for reproducibility, transparency, and benchmarking
  • conceived by Google in 2018
  • include key details:
    • the model type
    • the training parameters and experimental info (MLFlow is good for experiment tracking)
    • which datasets were used to train your model
    • the model’s evaluation results
  • should also provide a comprehensive assessment of a model’s intended usage, limitations, risks and mitigations, and ethical and safety considerations

https://modelcards.withgoogle.com/about

Example: ChatGPT-o1 System Card

https://openai.com/index/openai-o1-system-card/

Computing Metrics in Python

• You'll probably use skikit-learn for Machine Learning
• Depending on your industry, you'll need to report any of these evaluation results
• We practiced computing these by hand, because scikit-learn doesn't actually give you all the metrics you may need
  • scikit-learn doesn't compute FPR for you (is this true? https://machinelearningmastery.com/roc-curves-and-precision-recall-curves-for-classification-in-python/

https://scikit-learn.org/1.5/auto_examples/model_selection/plot_cost_sensitive_learning.html#sphx-glr-auto-examples-model-selection-plot-cost-sensitive-learning-py

Why we messing with all these metrics?

Stay tuned: these are so important for when we talk about fairness in AI

Public Service Announcement

Any time someone doesn't report at least precision and recall, but ideally a confusion matrix, it's as good as....

https://stats.stackexchange.com/questions/423/what-is-your-favorite-data-analysis-cartoon?page=2&tab=votes#tab-top