T-06: Applications of Machine Learning

01 Learning Objectives

By the end of this tutorial you will be able to

Identify the six broad capability areas where machine learning is most heavily deployed today.
Recognise which learning paradigm (supervised classification, regression, clustering, and so on) dominates each area.
Describe the recurring structural pattern that appears across very different application domains.
Explain why the quality of the training data is usually the dominant factor in whether an application succeeds.
Evaluate a claim about an ML application by asking what data it needs, what paradigm it uses, and how it would be evaluated.

02 Motivating Example

Consider a single ordinary morning for someone with a modern smartphone. Between waking up and arriving at work, they might interact with a dozen machine learning systems without noticing any of them.

They unlock the phone with their face — a face recognition model.
They dictate a message using voice — a speech recognition model.
They type a search query, and the results are ranked — a ranking model.
They read a news feed where the ordering is personalised — a recommendation model.
They open a map that predicts traffic conditions — a forecasting model.
Their bank declines a suspicious transaction — a fraud detection model.
A notification arrives suggesting they leave early for an appointment — a predictive scheduling model.
They take a photo and the phone automatically improves the lighting — an image enhancement model.

Not one of these systems shares data, model architecture, or evaluation criteria with any of the others. Face recognition operates on images; fraud detection operates on transaction records. Speech recognition processes audio; traffic prediction processes time series. Yet behind all of them sit the same four ingredients you met in T-05: features, labels, parameters, and loss.

The purpose of this tutorial is to develop a feel for where machine learning is applied, and to see the same underlying structure through the surface variety of very different applications.

03 Concept Explanation

Six capability areas

Applications of machine learning can be grouped into six broad capability areas. The grouping is not perfectly clean — some systems span more than one — but it captures most of what the field actually does in practice.

Perception: Turning raw sensory data — images, audio, video — into structured understanding. This is where deep learning dominates, and it is where most of the field's public success stories come from.
Prediction: Forecasting a value or outcome from structured data. Often called “tabular prediction” because the data is typically arranged in tables. Credit scoring, demand forecasting, and risk assessment all live here. Classical ML algorithms — not deep learning — tend to dominate this area.
Language: Processing and producing text. Translation, summarisation, sentiment analysis, and conversational assistants all fall into this category. Deep learning has transformed language processing over the past decade.
Ranking: Ordering a set of items by predicted relevance to a user. Search engines, recommendation systems, advertising platforms, and social media feeds are all fundamentally ranking problems. This is one of the most commercially valuable areas of machine learning.
Detection: Identifying rare or unusual events: fraud, network intrusions, equipment failures, spam. Detection problems have a distinctive character because the events of interest are almost always rare, and the cost of a false positive is often very different from the cost of a false negative.
Generation: Creating new content: images, text, code, audio. This is the newest of the six capability areas, and it is the one that has driven the most public attention to machine learning since 2022.

Which paradigm dominates where

Different capability areas tend to be dominated by different learning paradigms — the categories from T-05 — because the structure of the problem dictates which paradigm applies.

Supervised classification dominates perception, language, and detection. These problems involve predicting a category: what object is in this image? Is this email spam? Is this transaction fraudulent?
Supervised regression dominates prediction. These problems involve predicting a continuous value: what will the house sell for? How much demand will there be next month?
Ranking problems are a specialised form of supervised learning. The model is not asked to produce a single label or value; it is asked to produce a relative ordering.
Unsupervised learning is used less often in production than the other paradigms, but it appears in data exploration, customer segmentation, and anomaly detection — the subject of Module 4.
Generation uses a class of models called generative models, which learn to produce new examples that resemble the training data. These are usually deep neural networks, and they are covered in Module 5.

The recurring structural pattern

Look past the surface differences and almost every machine learning application follows the same structure:

There is a lot of historical data. The application succeeds because thousands or millions of past examples exist, each with a known outcome.
The pattern is stable enough to learn. The relationship between the input and the outcome is not random, and it does not change faster than the model can be retrained.
The decision is repeatable at scale. The same kind of decision is made many times, so automating it has value. If the decision is made once a year, machine learning is rarely worth the effort.
Being approximately right has value. A fraud detector that catches 95% of fraud cases is worth deploying, even though it is wrong 5% of the time. If the system had to be perfect, ML would not be the right approach.

This pattern explains why some applications have been transformed by machine learning and others have barely been touched. Where all four conditions hold — large historical datasets, stable patterns, high-volume decisions, tolerance for occasional error — the technology tends to succeed. Where any of them fails, it tends to disappoint.

Why data quality dominates everything else

An observation that surprises people who are new to the field: the choice of algorithm usually matters much less than the quality of the training data.

A well-curated dataset of a hundred thousand clean, correctly labelled examples will almost always produce a better model than a poorly curated dataset of a million examples, regardless of which algorithm is used. This is why machine learning projects spend the overwhelming majority of their time on data collection, cleaning, and labelling — not on modelling.

It is also why some applications are dramatically harder than they look from the outside. Medical diagnosis from X-rays sounds like a natural fit for machine learning — and it is — but obtaining a large dataset of correctly labelled X-rays requires radiologists to spend thousands of hours annotating images. Fraud detection sounds straightforward until you realise that only a tiny fraction of transactions are fraudulent, so the model has very few positive examples to learn from. Translation sounds difficult and is, but decades of parallel text — documents available in multiple languages — provided the training data almost for free.

Not every application succeeds

It is worth being honest that a large fraction of machine learning projects do not deliver what they promised. Industry surveys consistently find that a majority of ML projects fail to reach production, and the reasons are usually not technical: insufficient data, unclear objectives, inadequate integration with existing systems, or a mismatch between what the model predicts and what the business actually needs.

T-07 examines these failure modes in detail, and T-08 introduces the lifecycle that well-run projects follow to reduce the chance of joining that statistic. For now, the important point is that the capability areas described above are where the field has demonstrably worked — not a comprehensive list of where it has been attempted.

04 Visual Illustration

The diagram below maps the six capability areas and gives representative applications for each. Read it as a survey of the field rather than as an exhaustive list: most of what you will encounter as a practitioner falls somewhere within these six boxes.

Figure 1. The six capability areas where machine learning is most heavily deployed. Perception, Language, and Generation are dominated by deep learning; Prediction and Ranking use a wider mix, including classical ML; Detection problems appear across both paradigms but are usually defined by the rarity of the events of interest.

Reading the map

The six areas are not mutually exclusive. A self-driving car uses Perception (recognising pedestrians), Prediction (forecasting the movement of other vehicles), and Detection (identifying unusual situations). A large language model can be used for Language, Generation, and Ranking tasks simultaneously. The map is a useful starting point for orienting a new problem, not a rigid taxonomy.

05 Worked Example

Let us take a single application — credit card fraud detection — and work through it carefully. Fraud detection is a good case study because it appears in the news frequently, it is often misunderstood, and it illustrates several of the recurring patterns from this tutorial in a compact form.

Credit card fraud detection: a case study

The problem. A bank processes millions of card transactions every day. A small fraction of them — typically well under 1% — are fraudulent. The bank wants to identify fraudulent transactions in real time, before the money leaves the customer's account.

Step 1: Is this a good fit for machine learning? Apply the four conditions from the concept explanation:

Historical data? Yes — the bank has years of records, including transactions that were later confirmed as fraudulent.
Stable patterns? Partly. Fraud patterns change as fraudsters adapt, which means the model needs regular retraining. But short-term patterns are stable enough to learn.
High-volume decisions? Yes — millions of transactions per day.
Tolerance for error? Partly. A missed fraud is expensive; a false alarm blocks a legitimate purchase and annoys the customer. The bank has to balance these costs.

The application passes the test, but the fourth condition is where the difficulty lies.

Step 2: What is the task? This is a supervised classification problem. Each transaction is classified as fraudulent or legitimate. The capability area is Detection.

Step 3: What are the features and labels?

Features might include: transaction amount, merchant category, time of day, location, distance from the customer's home, whether the card was physically present, and the customer's recent transaction history.
Label is whether the transaction was actually fraudulent — determined, in the training data, by looking at what happened afterwards: did the customer report it? Was it refunded?

Step 4: What are the parameters and the loss? The parameters depend on the model chosen — a logistic regression would have one weight per feature, a gradient boosted tree would have hundreds of split points, a neural network would have thousands of weights. The loss function needs to be designed carefully: because fraud is so rare, a naive loss that just counts errors would lead the model to simply predict “legitimate” for every transaction, which would be correct 99.9% of the time. Instead, the loss must weight the two kinds of error differently.

Step 5: Why is this harder than it looks? Several complications make fraud detection genuinely difficult:

Extreme class imbalance. Fraudulent transactions are perhaps 0.1% of all transactions. The model sees very few positive examples, which makes learning the pattern harder.
Asymmetric costs. A missed fraud might cost hundreds of pounds; a false alarm might cost a customer relationship. The model must be tuned to balance these, and the balance point depends on business decisions, not statistical ones.
Adversarial environment. Unlike weather prediction, where the underlying process does not care about the forecast, fraud detection operates against opponents who actively try to evade it. Once a fraudster learns what the model flags, they adapt.
Distribution shift. The customer base changes, spending patterns change (holiday seasons, economic conditions), and fraud tactics change. A model trained on last year's data may perform poorly on this year's transactions.

Step 6: How would it be evaluated? Not with accuracy. A model that predicts “legitimate” for every transaction would achieve 99.9% accuracy while catching no fraud at all. Instead, the bank would use metrics like precision (of the transactions flagged as fraud, how many actually were?), recall (of the actual frauds, how many did the model catch?), and the trade-off between them. You will meet these metrics in Module 2, Unit 6.

This case study illustrates several themes from the tutorial. It is a Detection application, it uses supervised classification, and it satisfies the four conditions for a good ML fit — but only just. The extreme class imbalance, the adversarial nature of the problem, and the asymmetric costs of different errors are all features that make fraud detection substantially harder than it first appears. The same is true of most real-world applications: the interesting difficulty is almost never in the algorithm.

06 Interactive Activity

Classify the application

For each of the following applications, identify two things:

Capability area — which of the six boxes from Figure 1 does it fall into? (Perception, Prediction, Language, Ranking, Detection, Generation)
Learning paradigm — is it supervised classification, supervised regression, unsupervised learning, or something else?

Ten applications — identify capability area and learning paradigm
#	Application	Capability	Paradigm
1	A system that reads handwritten postal addresses and routes mail automatically.
2	A tool that estimates a property's market value from its size, location, and condition.
3	A service that flags unusual login attempts as potential account compromises.
4	A model that translates product descriptions from English into Spanish.
5	A system that suggests which articles a reader is most likely to click on.
6	A tool that generates realistic-looking portraits of people who do not exist.
7	A model that groups customers into segments based on their spending behaviour, without being told what the segments should be.
8	A system that predicts how much electricity a neighbourhood will consume tomorrow.
9	A tool that transcribes voicemail messages into text.
10	A system that ranks search results by predicted relevance to the query.

Show suggested answers

1. Postal address reading — Perception; supervised classification. The input is an image, the output is a sequence of digit or letter classifications.

2. Property valuation — Prediction; supervised regression. The output is a continuous value (a price), and the input is tabular data. This is a classic regression task.

3. Login anomaly detection — Detection; could be either supervised or unsupervised. If the bank has labelled examples of past compromises, it is supervised classification. If not, the system may use unsupervised anomaly detection to flag logins that look unusual relative to the user's normal pattern. In practice, most production systems use both.

4. Translation — Language; supervised learning. Translation is a sequence-to-sequence task, which is a specialised form of supervised learning. Historically the paradigm was statistical machine translation; today it is almost always deep learning.

5. Article recommendation — Ranking; supervised learning. The model is trained on which articles users actually clicked, and its job is to produce a relative ordering rather than a single prediction.

6. Portrait generation — Generation; generative modelling. The model learns the distribution of faces in a training set and can sample new ones. This is neither classification nor regression in the usual sense.

7. Customer segmentation — this cuts across Prediction and Detection; unsupervised learning. The task is grouping, not predicting. This is the domain of clustering, covered in Module 4.

8. Electricity demand forecasting — Prediction; supervised regression. Output is a continuous quantity (kilowatt-hours), input is historical time-series data. Classic regression problem.

9. Voicemail transcription — Perception and Language; supervised learning. This application genuinely spans two capability areas. The input is audio (Perception) and the output is text (Language). Modern systems handle it end-to-end with a single deep neural network.

10. Search result ranking — Ranking; supervised learning. Trained on click data and relevance labels, producing a relative ordering of results rather than a single prediction.

Several of these do not fit neatly into one box. This is normal. The capability map is a starting point for orienting a problem, not a rigid classification scheme.

07 Self-Check Questions

Try each question before revealing the answer. If you get one wrong, re-read the section it refers to before continuing.

Question 1

Which of the following is not one of the four conditions that tend to predict success for a machine learning application?

A large amount of historical data with known outcomes.
Patterns that are stable enough to learn.
Access to the latest neural network architectures.
Tolerance for occasional errors, since the system need not be perfect.

Show answer

C. The choice of architecture matters far less than the quality and quantity of the data. The other three conditions — data, stable patterns, and tolerance for approximation — are the ones that determine whether an application is likely to succeed.

Question 2

A model trained to detect fraudulent transactions achieves 99.9% accuracy on the test set. Why is this number not necessarily impressive?

Because test sets are always easier than real data.
Because fraud is rare — a model that predicts “legitimate” for every transaction would also achieve 99.9% accuracy, while catching no fraud at all.
Because accuracy is not a real machine learning metric.
Because fraud detection models are always trained on clean data.

Show answer

B. When one class is far more common than the other, accuracy becomes a misleading metric. A trivially useless model that always predicts the majority class can achieve very high accuracy. This is why detection tasks typically use precision, recall, and related metrics instead. You will meet these in Module 2, Unit 6.

Question 3

Which of the six capability areas is most dominated by deep learning?

Prediction — forecasting values from structured data.
Perception — understanding images, audio, and other sensory data.
Ranking — ordering items by relevance.
Detection — finding rare events.

Show answer

B. Perception problems involve high-dimensional, unstructured input (pixels, audio waveforms) where deep learning has demonstrated a decisive advantage. Prediction and Ranking problems often work well with classical ML, and Detection problems use a mix of paradigms depending on the data available.

Question 4

Fraud detection is often described as harder than weather prediction, even though both are supervised learning problems. What is the main reason?

Fraud detection involves more data.
Fraudsters actively adapt to evade the model, making the pattern adversarial and shifting over time.
Fraud detection requires deep learning, whereas weather prediction does not.
Fraud detection is not actually harder — the two problems are equivalent.

Show answer

B. Fraud detection operates in an adversarial environment: the entities the model is trying to detect actively change their behaviour in response. Weather does not care about the forecast. This makes the underlying data distribution shift over time, which means fraud models need frequent retraining and constant monitoring.

Question 5

A company announces that it will use machine learning to decide which employees to promote, based on historical promotion data. Why is this application likely to be problematic, even if the model is technically well-built?

Because promotion decisions are not suitable for machine learning.
Because the historical data likely reflects past biases, which the model will learn and reproduce.
Because the company does not have enough Computing power.
Because promotions cannot be predicted from any kind of data.

Show answer

B. Machine learning learns from historical patterns — including historical patterns of discrimination. If past promotion decisions disadvantaged certain groups, a model trained on that data will reproduce and amplify those patterns. This is one of the central ethical issues in applied machine learning, and it is the subject of Module 6. The technical machinery works fine; the problem is what it is being asked to do.

08 Practice Exercises

Classify the capability area. For each application, state which of the six capability areas from Figure 1 it belongs to (Perception, Prediction, Language, Ranking, Detection, Generation).
- Converting scanned handwritten documents into editable text.
- Predicting tomorrow's stock price for a given company.
- Deciding which customer support tickets should be handled first.
- Flagging unusual activity in a user's account.
- Producing a summary of a long news article.
- Suggesting which products a user might want to buy.
Classify the paradigm. For each of the following, state whether it uses supervised classification, supervised regression, unsupervised learning, or something else.
- Predicting whether a loan applicant will default.
- Grouping news articles by topic without predefined categories.
- Estimating the price of a used car.
- Translating a sentence from English to French.
Identify the four conditions. For each application, state which of the four success conditions (large historical data, stable patterns, high-volume decisions, tolerance for approximation) might fail.
- Predicting earthquakes.
- Diagnosing rare genetic diseases.
- Recommending films to users.
- Detecting fraudulent insurance claims.
True or false, with justification.
- All machine learning applications involve prediction.
- Ranking is a form of supervised learning.
- The choice of algorithm is the most important factor in whether a machine learning application succeeds.

09 Problem Sets

Analytical problems

Consider two applications: automatic speech recognition and predicting customer churn. Both use supervised learning, but they differ in the structure of the input data, the availability of training data, and the cost of errors.

(a) For each application, describe the features, labels, and loss. (b) Which application has more training data readily available, and why? (c) Which application is more sensitive to errors of different kinds, and how would this affect the design of the system?
The tutorial claims that data quality dominates algorithm choice. Give three specific scenarios in which this claim might not hold, i.e., where the choice of algorithm matters more than the quality of the data.
A team is building a system to detect fraudulent transactions. The fraud rate in their data is 0.1%. A naive model that predicts “not fraud” for every transaction achieves 99.9% accuracy.

(a) What metric should they use instead of accuracy? (b) What are the practical consequences of using accuracy as the evaluation metric in this setting? (c) How might the imbalance itself affect the training process, not just the evaluation?
Some applications of machine learning are described as “adversarial” — that is, the entities the model is trying to detect actively change their behaviour in response. Give three examples beyond fraud detection, and for each one describe a specific way the adversarial nature of the problem affects the design of the system.

Technical problems

The following table describes three machine learning systems. Complete it by filling in the missing entries.

Application	Capability area	Paradigm	Most likely approach
Predicting tomorrow's temperature
Classifying images of skin lesions
Grouping customers by purchasing habits
Ranking web pages for a search query
Generating synthetic speech

A fraud detection dataset contains 1,000,000 transactions, of which 500 are fraudulent. A model trained on this data produces the following confusion matrix on a held-out test set:

	Predicted: Fraud	Predicted: Legitimate
Actual: Fraud	320	180
Actual: Legitimate	900	998,600

(a) Compute the accuracy. (b) Compute the precision of the fraud predictions. (c) Compute the recall of the fraud predictions. (d) Why is the accuracy a misleading metric here? (e) In this application, which error type is more costly — false positives or false negatives? Justify your answer.

A team is building a recommendation system for a streaming platform with 10 million users and 50,000 titles. The system must produce recommendations for each user within 100 milliseconds.

(a) What capability area is this? (b) Why can the system not simply compute a full recommendation for every user and every title on every request? (c) Describe a plausible two-stage architecture that would allow the system to meet the latency constraint. (Hint: consider that the first stage might produce a candidate set from which the second stage selects.)

Theoretical questions

The tutorial identifies six capability areas. Are these mutually exclusive? Could you imagine a seventh capability area that does not fit into any of the six? If so, what would it be, and what distinguishes it?
Detection is described as fundamentally harder than other applications because the events of interest are rare. Is rarity the only thing that makes detection difficult, or are there other structural features that characterise detection problems?
The tutorial claims that the four success conditions predict whether an application will succeed. Are any of these conditions more important than the others? Is it possible for an application to succeed while failing one of the conditions?
Ranking problems are described as a specialised form of supervised learning. What specifically distinguishes ranking from classification or regression, and why does that distinction matter for how the model is trained?
The tutorial notes that a large fraction of ML projects fail to reach production, and that the reasons are usually not technical. What does this suggest about the skills that machine learning practitioners need, beyond their technical training?

10 Discussion and Projects

Discussion questions

Which of the six capability areas do you think will see the most transformative applications over the next decade? Justify your prediction with reference to the four success conditions.
Some applications of machine learning are described as “solved” — for example, image classification on standard benchmarks. What does it mean for a problem to be solved? Are there solved problems where deployment still fails in practice?
The tutorial argues that data quality dominates algorithm choice. Does this suggest that machine learning practitioners should spend less time learning about algorithms? What is the right balance?
Some applications of machine learning are ethically contested, such as predictive policing or automated hiring. Where should the line be drawn between applications that should be built and those that should not? Who decides?

Case study: AlphaFold and protein structure prediction

Case study

Proteins are chains of amino acids that fold into complex three-dimensional shapes. The shape of a protein determines its function, and understanding that shape is essential for drug discovery, understanding diseases, and designing new enzymes. But determining a protein's shape experimentally — through techniques like X-ray crystallography or cryo-electron microscopy — is slow and expensive. For decades, predicting a protein's shape directly from its amino acid sequence had been one of the great open problems in biology.

In 2020, DeepMind's AlphaFold2 system achieved a breakthrough in the Critical Assessment of Structure Prediction (CASP) competition. The system's predictions were, in many cases, accurate to within the margin of experimental error — a result that many researchers had expected to take another decade to achieve.

The system was trained on the Protein Data Bank, a public database of experimentally determined protein structures that has been accumulating for over fifty years. It used deep learning, specifically a type of architecture called a transformer, and consumed significant computational resources during training.

Analysis questions:

Which capability area and which paradigm does AlphaFold fall into?
Which of the four success conditions does AlphaFold satisfy? Are there any it does not satisfy?
The Protein Data Bank has been accumulating for fifty years. What role did this decades-long data-collection effort play in AlphaFold's success?
AlphaFold's predictions are sometimes wrong. In the context of drug discovery, what would be the consequences of acting on a wrong prediction? What safeguards should be in place?
Is AlphaFold's success primarily a triumph of algorithmic innovation, of data availability, or of compute? Justify your answer.

Critical-thinking activity: evaluating a machine learning business proposal

Activity

A startup pitches the following business to a venture capital firm:

“We're building a machine learning system to predict which students will drop out of university in their first year, so that universities can intervene early. We have initial data from three universities and are seeking funding to expand to fifty.”

In groups, prepare a due diligence report on this proposal. Address:

Which capability area and paradigm would this system use?
Which of the four success conditions are likely to hold, and which might not?
What are the failure modes from T-07 that this system is most vulnerable to?
What questions would you want answered before investing?
What ethical concerns would you want the founders to address?
Would you invest? Why or why not?

The purpose is to practise applying the analytical framework from this tutorial to an unfamiliar application, without the benefit of seeing how it turns out.

Homework assignment

Assignment

Choose one of the six capability areas (Perception, Prediction, Language, Ranking, Detection, Generation). Write a 1,500-word survey of how machine learning is applied within that area.

Your survey should:

Define the capability area precisely.
Describe at least four different applications within the area.
Identify which learning paradigms dominate the area and why.
Discuss at least two failure modes that are particularly relevant to the area.
Analyse one application in depth, including what data it uses and how it is evaluated.
Predict one direction in which the area is likely to develop.

Cite at least five sources, at least two of which should be technical (research papers or engineering blog posts from the teams building these systems).

Mini-project: catalogue and evaluate ML applications

Mini-project

Build a catalogue of twenty real machine learning applications, five from each of four capability areas of your choosing (or spread across all six). For each application, record:

The name of the system and the organisation behind it.
The capability area and paradigm.
The likely features and labels.
At least one failure mode it is vulnerable to.
One interesting or unusual aspect of the application.

Then write a 1,000-word analysis of your catalogue, addressing:

Which capability areas were easiest to find examples for? Which were hardest?
What patterns did you notice across the applications?
Which applications surprised you, and why?
Are there capability areas or application types that seem conspicuously absent from the public discourse? Why might that be?

The catalogue can be built as a spreadsheet, a document, or a web page. The important thing is the curation and the analysis that follows.

11 Summary

Key takeaways from this tutorial

Six capability areas capture most of what machine learning does in practice: Perception, Prediction, Language, Ranking, Detection, and Generation.
Different areas favour different paradigms. Perception, Language, and Detection lean toward supervised classification. Prediction uses regression. Ranking is its own specialised form. Generation uses generative models.
Four conditions predict success. Large historical datasets, stable patterns, high-volume decisions, and tolerance for occasional error. When all four hold, applications tend to work; when any fails, they tend to disappoint.
Data quality dominates algorithm choice. A well-curated dataset with a simple model almost always outperforms a poorly curated dataset with an advanced model.
Real applications are harder than they look. Class imbalance, adversarial environments, and asymmetric costs all complicate deployment in ways that the textbook algorithms do not anticipate.
Some failures are ethical, not technical. A model that accurately reproduces historical discrimination is working exactly as designed — which is precisely the problem.

12 Further Reading

The following sources expand on the applications surveyed in this tutorial.

Jordan, M. I., & Mitchell, T. M. (2015). “Machine learning: Trends, perspectives, and prospects.” Science, 349(6245), 255–260.
A broad survey of machine learning applications written for a general scientific audience. Useful for seeing the landscape beyond the applications covered here.
Sculley, D., et al. (2015). “Hidden Technical Debt in Machine Learning Systems.” Advances in Neural Information Processing Systems, 28.
A short paper that has become required reading in the field. It describes why machine learning systems are harder to build and maintain than they appear from the outside, and it pairs well with the section on failure modes in this tutorial.
Esteva, A., et al. (2017). “Dermatologist-level classification of skin cancer with deep neural networks.” Nature, 542(7639), 115–118.
A concrete medical application paper. Read it as a case study in how a machine learning application is actually built, evaluated, and defended — not just as an example of deep learning working well.
O'Neil, C. (2016). Weapons of Math Destruction. Crown.
A critical examination of how machine learning models are deployed in high-stakes decisions — hiring, lending, criminal justice — and where they cause harm. Not a technical book, but an important one. It pairs with Module 6.

Next up

T-07 examines the failure modes of machine learning in more depth. Many of the themes touched on in this tutorial — class imbalance, adversarial environments, biased historical data — are examined there from the perspective of what goes wrong and why.