T-09: Unit Review and Self-Assessment

01 Learning Objectives

By the end of this tutorial you will be able to

Consolidate the eight tutorials of Unit 1 into a coherent picture of machine learning's purpose, history, mechanics, and practice.
Trace the dependencies between the concepts you have encountered — which ideas depend on which others.
Diagnose your own understanding by completing the unit quiz and interpreting your score.
Identify which tutorial to revisit if your quiz performance reveals a gap.
Prepare for Unit 2 by recognising what you will need to carry forward from this unit.

This is not new content

T-09 does not introduce any new concepts. Everything here is a reminder or a re-framing of material from T-01 through T-08. Read it quickly, then spend your real effort on the quiz at the end. The quiz is the point.

02 Why Review Matters

There is a well-documented phenomenon in learning research called the testing effect. When students study material and then take a test on it, they retain substantially more of it weeks later than students who study the same material for the same amount of time but do not take a test.

The effect is counterintuitive. Re-reading a tutorial feels productive. You recognise the material, you feel confident you understand it, and you close the page with a sense of completion. Taking a quiz feels uncomfortable — you might get questions wrong, which is unpleasant. By any intuitive measure, re-reading should be better.

But recognition is not the same as recall. Re-reading builds familiarity. Quizzing builds the ability to retrieve information from memory without prompts, which is what actually matters when you encounter a problem in a new context. This is why the field of machine learning will seem different to you after you have used it than it does now, when you have only read about it.

The quiz at the end of this tutorial is not a test of whether you deserve to move on. It is a diagnostic: a way of finding out what you actually retained, and what you only recognised. If you get questions wrong, the correct response is not frustration but attention — you have just found the specific material worth revisiting.

How to use this tutorial

Skim sections 03 through 06 for the review. Then, when you reach the quiz in section 07, close the other tutorials and take it cold. Score it honestly. Only then consult the answers, and use the links in each answer to return to the relevant tutorial if you need to.

03 Concept Review

Unit 1 covered eight tutorials. They cluster into four themes, and it is worth reading through these summaries in order — each theme builds on the one before.

Theme 1: What machine learning is

T-01 gave the central definition: machine learning is the practice of building systems that improve at a task through experience, rather than through explicitly programmed rules. Tom Mitchell’s formal framing — task T, experience E, performance measure P — captures the same idea. The reversal at the heart of the field is that classical programming takes rules and produces answers, whereas machine learning takes data and answers and produces rules.

T-04 clarified the vocabulary. Artificial intelligence is the broadest term, machine learning is a subset of AI, and deep learning is a subset of machine learning. The exception that the tidy nested diagram omits is symbolic AI — rule engines and compilers that are AI but not ML.

Key terms: task, experience, performance measure, feature, label, supervised learning, unsupervised learning, symbolic AI, hybrid systems.

Theme 2: Where machine learning came from

T-02 traced the field’s history through seven eras. Two AI winters — roughly 1974–1980 and 1987–1993 — followed periods of unmet promises. The current deep learning era began in 2012, when AlexNet’s victory in the ImageNet competition made it clear that the combination of large datasets, GPU compute, and algorithmic refinements had finally made neural networks practical.

T-03 examined the shift from symbolic to statistical methods that took place through the 1990s. Symbolic AI encoded knowledge as hand-written rules; statistical learning extracts patterns from data. The symbolic approach failed not because its rules were wrong but because of the knowledge acquisition bottleneck: experts cannot always articulate what they know, and encoding their knowledge as rules is slow and brittle.

Key terms: AI winter, symbolic AI, statistical learning, knowledge acquisition bottleneck, brittleness, probabilistic output.

Theme 3: How machine learning works

T-05 introduced the mechanics that underlie every algorithm in the course. Four terms: features and labels come from the data, parameters are what the algorithm learns, and loss is a single number measuring how wrong the model currently is.

The training loop has four steps: predict, measure the loss, adjust the parameters, repeat. Every algorithm — linear regression, decision trees, neural networks — runs this loop. The differences between algorithms are in how the model is represented, how the loss is defined, and how the parameters are adjusted, but the loop itself is always the same.

Key terms: feature, label, parameter, loss, prediction, training loop, learning rate, generalisation.

Theme 4: How machine learning is used, and how it fails

T-06 surveyed the six capability areas where machine learning is deployed: Perception, Prediction, Language, Ranking, Detection, and Generation. It identified four conditions that predict success — large historical datasets, stable patterns, high-volume decisions, and tolerance for approximation — and noted that data quality dominates algorithm choice in determining whether a project will work.

T-07 examined seven failure modes: overfitting, underfitting, spurious correlation, data leakage, distribution shift, class imbalance, and feedback loops. The tutorial’s central observation is that none of these is fundamentally an algorithm problem. They are problems with data, deployment, or assumptions. This is why practical machine learning spends most of its time on data work and monitoring.

T-08 introduced the seven-stage lifecycle that formalises this: problem framing, data collection, data cleaning, feature engineering, modelling, evaluation, and deployment and monitoring. The lifecycle is a loop rather than a line, and the distribution of effort is heavily skewed toward data work — roughly 60–70% of a project's time — with modelling and evaluation together accounting for only 10–20%.

Key terms: capability area, overfitting, underfitting, bias–variance tradeoff, spurious correlation, data leakage, data drift, concept drift, class imbalance, feedback loop, lifecycle stages.

04 Concept Map

The diagram below shows how the eight tutorials of Unit 1 relate to one another. The arrows indicate dependencies: an arrow from one tutorial to another means the second tutorial assumes the first. If you found any tutorial difficult, the arrows pointing into it will show you which earlier material is worth revisiting.

Figure 1. The dependency structure of Unit 1. T-05, the mechanics tutorial, is the pivot point: it depends on the foundational tutorials above it and enables all the practical tutorials below it. If you are going to re-read one tutorial before the quiz, T-05 is the one whose absence would cause the most difficulty downstream.

05 Worked Synthesis

The quiz in the next section tests recall of specific concepts. This worked example tests something harder: whether you can combine ideas from several tutorials to reason about a new situation. Read it carefully — it models the kind of thinking that the rest of the course will ask you to do.

A synthesis scenario

A hospital wants to build a system that identifies patients at high risk of developing sepsis within the next 48 hours. Sepsis is a life-threatening response to infection, and early detection improves survival rates significantly. A team of clinical researchers and engineers is assembled to build the system.

Working through the synthesis, using ideas from across Unit 1:

What kind of problem is this? It is a supervised classification problem (T-05) — the output is a binary decision, and the training data will consist of historical patients with known outcomes. It falls in the Detection capability area (T-06), because sepsis is rare and the interesting cases are the ones the model needs to catch.
Is it a good fit for machine learning? Applying the four conditions from T-06: there is historical data (patient records with outcomes); the patterns are reasonably stable (sepsis has consistent biological signatures); the decision is high-volume (every admitted patient must be assessed); and there is tolerance for approximation (predictions are advisory, not diagnostic). The application satisfies the conditions.
What failure modes are likely? Several from T-07 apply. Class imbalance, because sepsis is rare. Data leakage, if any features are only recorded after a diagnosis is suspected. Distribution shift, because patient populations change and clinical practices evolve. Spurious correlation, if some features are correlated with sepsis for incidental reasons — for example, if one hospital unit routinely runs more tests than another and therefore catches more cases.
What does the lifecycle look like? Problem framing (T-08) comes first and is unusually consequential here: what exactly counts as “sepsis”? What horizon? What is the intervention the prediction triggers? Data collection involves extracting features from electronic health records, which are notoriously messy. Cleaning must handle missing values, inconsistent units, and time-zone issues. Feature engineering is where clinical expertise matters most. Evaluation cannot use accuracy — a model that predicts “no sepsis” for every patient would achieve very high accuracy while being useless.
Where on the symbolic-statistical spectrum does this sit? From T-03 and T-04: the system is primarily statistical, but a good one will include symbolic elements. Clinical rules that must always be respected (for example, alerting on any patient with a temperature above a certain threshold, regardless of the model's prediction) are symbolic overrides. The model is not deployed alone; it is deployed alongside the rules the clinical team already uses.
What are the ethical considerations? From T-06 and T-07: the historical data on which the model trains reflects historical clinical practice, which may have been uneven across demographic groups. If, for example, one patient group received less thorough testing historically, the model will learn to under-detect sepsis in that group. This is a feedback loop in the sense of T-07. Deploying a model that systematically underperforms for a specific population would be both unjust and dangerous, and it would arguably violate medical standards.

Notice what this synthesis required. No single tutorial was sufficient. T-05 gave the mechanics, T-06 gave the framing conditions, T-07 gave the failure modes, T-08 gave the lifecycle, T-03 and T-04 gave the symbolic-statistical context, and T-01 gave the underlying definition. The whole unit was necessary to reason about one project carefully.

This is the point of the review. Unit 1 is not a collection of eight unrelated topics — it is a single integrated toolkit. The quiz that follows checks whether you have the individual pieces. The capstone project and every assignment in the course will check whether you can assemble them the way this worked example does.

06 Confidence Audit

Rate your understanding before the quiz

Before taking the quiz, spend a few minutes on this audit. For each concept below, rate your confidence on a scale of 1 (not confident) to 5 (could explain it to someone else). Be honest — the point is to identify what to revisit, not to prove anything.

If your confidence is 1 or 2 on any item, note which tutorial covers it (listed in brackets) and revisit that tutorial before taking the quiz.

Rate your confidence: 1 (low) to 5 (high)
#	Concept	Tutorial
1	Mitchell's definition of machine learning	T-01
2	The difference between classical programming and machine learning	T-01
3	The two AI winters and what caused them	T-02
4	The knowledge acquisition bottleneck	T-03
5	The difference between symbolic AI and statistical learning	T-03
6	The nested relationship between AI, ML, and DL — including the symbolic exception	T-04
7	Features, labels, parameters, and loss	T-05
8	The four steps of the training loop	T-05
9	The six capability areas	T-06
10	The four conditions that predict ML success	T-06
11	Overfitting versus underfitting	T-07
12	Data leakage and how to spot it	T-07
13	Distribution shift: data drift versus concept drift	T-07
14	The seven stages of the ML lifecycle	T-08
15	Where the time actually goes in an ML project	T-08

Interpreting your ratings

If you rated everything 4 or 5, you may be overestimating your understanding — the quiz will tell you more honestly than your own sense of confidence. If several items are 1 or 2, that is not a problem. It means you know which tutorials to revisit before continuing, which is exactly what this audit is for.

07 Unit Quiz

Twelve questions covering all eight tutorials. Take it cold — close the other tutorials first. Score it honestly, and use the answers to decide whether there is anything to revisit before moving to Unit 2.

12

questions · ~20 minutes · diagnostic

Question 1

Which of the following best captures what is distinctive about machine learning compared with classical programming?

Machine learning uses computers; classical programming does not.
Machine learning takes rules and produces answers; classical programming takes answers and produces rules.
Machine learning takes data and answers, and produces rules; classical programming takes rules and data, and produces answers.
Machine learning requires more computational power than classical programming.

Show answer

C. The reversal — rules as output rather than input — is the defining shift. Option B reverses the two descriptions incorrectly. See T-01.

Question 2

In Mitchell’s definition, a program learns from experience E with respect to task T and performance measure P if its performance at T, as measured by P, improves with E. For a system that predicts whether a student will pass a course based on historical records of study habits and outcomes, what is the experience E?

The study habits of the student.
Predicting whether a student will pass.
The historical student records, including their outcomes.
The fraction of predictions that turn out to be correct.

Show answer

C. The experience is the training data — the historical records the system learns from. Option A is a feature; option B is the task; option D is the performance measure. See T-01.

Question 3

Which of the following is not a subset of machine learning?

Deep learning.
Convolutional neural networks.
Symbolic expert systems.
Random forests.

Show answer

C. Symbolic expert systems are AI but not ML — they operate on hand-written rules rather than learning from data. This is the exception to the neat nested hierarchy. See T-04.

Question 4

What is an “AI winter”?

A period when AI systems stop working due to hardware failures.
A period when funding and public interest in AI contract sharply following unmet expectations.
A period when AI research moves from one country to another.
A period when AI systems become too expensive to run.

Show answer

B. Two AI winters occurred, roughly 1974–1980 and 1987–1993, following periods of unmet promises. See T-02.

Question 5

Which of the following best describes the difference between symbolic AI and statistical learning?

Symbolic AI uses computers; statistical learning does not.
Symbolic AI encodes knowledge as hand-written rules; statistical learning extracts patterns from data.
Symbolic AI is faster than statistical learning.
Symbolic AI is used in industry; statistical learning is used in academia.

Show answer

B. The essential difference is where the knowledge comes from. Symbolic AI gets it from human experts writing rules; statistical learning gets it from data. See T-03.

Question 6

In a model that predicts house prices from floor area, number of bedrooms, and distance to the nearest school, which of the following is a parameter?

Floor area.
Sale price.
The weight assigned to floor area by the model.
The average sale price in the training data.

Show answer

C. The weight is a parameter — a number the algorithm learns. Options A and B are the feature and label respectively; option D is a statistic of the data, not a model parameter. See T-05.

Question 7

What is the correct order of steps in the training loop?

Adjust parameters → Predict → Measure loss → Repeat
Predict → Measure loss → Adjust parameters → Repeat
Measure loss → Predict → Adjust parameters → Repeat
Adjust parameters → Measure loss → Predict → Repeat

Show answer

B. The loop is: predict with current parameters, measure the loss, adjust the parameters, then repeat. See T-05.

Question 8

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

Prediction — forecasting values from tabular 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 where deep learning has demonstrated a decisive advantage. Prediction and Ranking often work well with classical ML. See T-06.

Question 9

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

Large historical datasets with known outcomes.
Patterns that are stable enough to learn.
Access to the latest neural network architectures.
Tolerance for occasional errors.

Show answer

C. The choice of architecture matters far less than the quality and quantity of the data. The other three conditions plus high-volume decisions are the ones that determine success. See T-06.

Question 10

A model reports very low error on the training data but high error on a held-out validation set. Which of the following is the most likely explanation?

Underfitting — the model is too simple.
Overfitting — the model has memorised the training data.
Class imbalance — one class is too rare.
The training set is too large.

Show answer

B. The gap between low training error and high validation error is the signature of overfitting. See T-07.

Question 11

A hospital readmission model is trained using a feature called “number of follow-up appointments scheduled.” During testing, the model achieves excellent accuracy. In deployment, it fails completely. What is the most likely reason?

Overfitting to the training data.
Data leakage — the feature is only populated after a patient has been discharged, so it would not be available at prediction time.
Distribution shift between the training and test data.
The model is too simple for the problem.

Show answer

B. This is the canonical data leakage example. The feature is not available at prediction time, but it was available during training, so the model learned to use information that will not exist in deployment. See T-07.

Question 12

A team discovers during evaluation that their model's performance is poor. According to the machine learning lifecycle, which stage should they generally return to first?

Problem framing, because the prediction may be wrong.
Feature engineering or modelling — the stages that most directly determine how well the model captures the pattern.
Deployment, because the model is not working in production.
Data cleaning, because the data must be the problem.

Show answer

B. Poor evaluation usually points to the features the model is using or to the model itself. Framing can also be the culprit, but feature engineering and modelling are the most common first answers. See T-08.

Interpreting your score

10–12 correct: You are ready for Unit 2. Move on.

7–9 correct: You have a solid grounding, but certain topics are worth a quick re-read. Look at which questions you missed and return to the corresponding tutorials.

Below 7 correct: Before moving to Unit 2, revisit the tutorials corresponding to the questions you missed. Unit 2 builds directly on this material — particularly on the concepts from T-05 (features, labels, parameters, loss) and T-07 (failure modes).

08 Practice Exercises

These exercises draw on material from the entire unit. They are designed to test whether the concepts from T-01 through T-08 have consolidated into a coherent understanding.

Cross-tutorial recall. Answer each of the following in one sentence.
- What are the four terms that underlie every machine learning problem?
- What are the four steps of the training loop?
- What are the two AI winters, and what triggered each one?
- What is the difference between AI, ML, and DL?
- What are the four conditions that predict ML application success?
- Name three of the seven failure modes.
- What are the eight stages of the lifecycle?

Match the tutorial. Each of the following terms was introduced in a specific tutorial. Identify which one.

Term	Tutorial
Knowledge acquisition bottleneck
Bias–variance tradeoff
Feature
Data leakage
Iterative lifecycle
Deep learning

Define the term. For each term, write a one-sentence definition in your own words.
- Machine learning
- Spurious correlation
- Concept drift
- Data leakage
- Overfitting

Complete the timeline. Fill in the missing entries.

Era	Years
Foundations
Symbolic AI
First AI winter
Expert systems
Second AI winter
Statistical turn
Deep learning

09 Problem Sets

Analytical problems

The following scenario involves concepts from multiple tutorials. Read it carefully and answer the questions below.

A startup is building a system to predict whether small businesses will default on their loans. They have collected data on 5,000 historical loans, of which 200 defaulted. They train a deep neural network on the data and achieve 96% accuracy on a held-out test set. However, when deployed, the system performs poorly, and its predictions are heavily skewed toward approving loans for businesses in the technology sector.

(a) Is this an AI, ML, or DL system? Justify your answer using the distinctions from T-04. (b) Identify the feature, label, parameter, and loss. (c) Which of the seven failure modes from T-07 might explain the poor deployment performance? (d) Where in the lifecycle (T-08) would the team need to return to fix the problem? (e) What additional information would you want before recommending a specific course of action?
"Most of what determines whether a machine learning project succeeds happens before any model is trained and after any model is deployed." This claim is made in T-08. Defend or challenge this claim using evidence from at least four tutorials in this unit.
The statistical turn of the 1990s (T-03) and the deep learning breakthrough of 2012 (T-02) are both described as shifts in the field. In what ways are these shifts similar? In what ways are they different? Is the deep learning era better understood as a continuation of the statistical turn or as a distinct shift?
The unit has emphasised that data quality dominates algorithm choice, and that most failures are not algorithm problems. Construct a scenario in which this general claim would not hold, i.e., where the choice of algorithm would matter more than data quality.
The husky/wolf classifier (T-07), the Amazon hiring algorithm (T-01), and the COMPAS risk assessment tool (T-07) are all cases of machine learning systems that failed. What common themes connect these failures? What distinguishes them? What do they suggest about the kinds of failures that are most likely in practice?

Technical problems

Design a machine learning system for the following problem: a small online retailer wants to predict which customers are likely to leave a negative review, so that its customer service team can reach out proactively.

Your design should specify, using the vocabulary of this unit:
- The capability area and paradigm (T-06).
- The features, label, parameters, and loss (T-05).
- The four success conditions from T-06, and whether each holds.
- The two most significant failure modes from T-07 that the system is vulnerable to.
- The lifecycle stages (T-08) that will require the most attention.

A team has built a model that predicts customer churn. When evaluating the model on the test set, they obtain the following confusion matrix:

	Predicted: Churn	Predicted: No churn
Actual: Churn	120	80
Actual: No churn	250	1,550

(a) Compute the accuracy. (b) Compute the precision and recall of the "churn" class. (c) If the retention team can contact 300 customers per week, which metric matters more: precision or recall? Explain. (d) What would you need to know before deciding whether this model should be deployed?

The following timeline describes events in a machine learning project. For each event, identify which stage of the lifecycle it belongs to, and identify any failure modes from T-07 that the event might signal.

#	Event	Stage	Potential failure
1	Team meets with business stakeholders to define what "at-risk customer" means.
2	Team discovers that one of the datasets has 15% missing values in a key field.
3	Team notices a feature that is strongly correlated with the target but only in the last month of data.
4	Model performs well on validation but poorly when tested on data from six months later.
5	Deployment team reports that model inference takes 3 seconds, which exceeds the 500ms budget.
6	Six weeks after deployment, monitoring shows a 20% drop in precision.

Theoretical questions

The unit has introduced a large number of concepts. Which concept do you think is most important for a working practitioner to internalise, and why? Which concept do you think is most often overlooked in practice?
The historical narrative in T-02 presents AI as progressing through waves of optimism and disappointment. Is this pattern inevitable, or is it a consequence of how AI research has been funded and organised? Could a different institutional structure produce smoother progress?
The lifecycle from T-08 is presented as a way to reduce risk. But the lifecycle itself makes certain assumptions about how projects are structured. What assumptions does it make, and how might they break down for research projects, solo practitioners, or very small teams?
The distinction between AI, ML, and DL (T-04) is often blurred in public discourse. To what extent does this blurring reflect a genuine feature of the technology, and to what extent is it a problem that should be fixed?
The unit has argued that failures are usually data problems, not algorithm problems. But this claim itself is a simplification. What would a more nuanced version of the claim look like? What would it add?
Symbolic AI (T-03) is described as having genuine strengths that remain relevant, even though it is no longer the dominant approach. If you were building a new system today, in what circumstances would you choose a symbolic approach over a statistical one?

10 Discussion and Projects

Discussion questions

Reflecting on the unit as a whole: what was the most surprising thing you learned? What idea do you think will be most useful to you in future work?
The unit has been primarily conceptual. Which concepts do you think will become clearer once you start working with actual data and models? Which do you expect to remain abstract?
The unit has emphasised the practical, engineering, and ethical dimensions of machine learning. Do you think this emphasis is appropriate for an introductory course, or should the course focus more on algorithms and mathematics?
Based on everything in Unit 1, what do you think the biggest open problem in machine learning is? Is it a technical problem, a data problem, or something else?
The unit has introduced several competing framings: symbolic vs statistical, supervised vs unsupervised, ML vs DL. Do you find these framings useful, or do they oversimplify a more complex reality?

Case study: integrating everything

Case study

A regional public health authority wants to build a system to predict which patients arriving at emergency departments are at risk of developing sepsis. The system will be used to prioritise patients for additional monitoring.

The authority has three years of emergency department data covering about 200,000 visits. Of these, roughly 3,000 patients developed sepsis during their stay. The data includes vital signs, lab results, patient demographics, and free-text notes written by triage nurses.

The authority is considering two proposals. Proposal A uses a gradient boosted tree on structured features, achieving an estimated sensitivity of 78% and specificity of 92%. Proposal B uses a hybrid system that combines the tree with a set of hand-written clinical rules developed by an experienced emergency physician, achieving an estimated sensitivity of 82% and specificity of 88%. Proposal B is more expensive to build and maintain.

Analysis questions:

Which capability area and paradigm does this system fall into? (T-06)
Identify the feature, label, parameter, and loss for Proposal A. (T-05)
Which of the four success conditions from T-06 hold, and which might not? (T-06)
Which failure modes from T-07 is this system most vulnerable to?
Is the hybrid approach of Proposal B a form of symbolic AI, statistical learning, or something else? Justify your answer using the concepts from T-03 and T-04.
How would you decide between the two proposals? What additional information would you want?
Map the recommended implementation onto the eight stages of the lifecycle (T-08). Which stages would require the most attention?
What ethical considerations would you raise with the health authority?

Critical-thinking activity: write your own exam questions

Activity

Working in pairs, write five exam questions that you think would fairly test whether a student had understood Unit 1. Your questions should:

Cover at least four different tutorials from the unit.
Include at least one question that requires applying a concept, not just recalling a definition.
Include at least one question that has no single correct answer, and would require the student to argue a position.
Include at least one question that tests whether the student can connect concepts from different tutorials.

For each question, write a short model answer. Then exchange your questions with another pair and answer theirs. Discuss what makes a question good and what makes it unfair or trivial.

The purpose is not to generate exam content but to practise thinking about what understanding actually looks like. Writing good questions is a form of analysis in itself.

Homework assignment

Assignment

Write a 2,000-word reflection essay on what you have learned in Unit 1. Your essay should:

Identify the three concepts from the unit that you consider most important, and explain why.
Identify one concept that you found difficult or counterintuitive, and explain what you now understand about it.
Connect at least three concepts from different tutorials, showing how they relate to one another.
Apply one of the frameworks from the unit (Mitchell's definition, the lifecycle, the seven failure modes, or the six capability areas) to a real system or problem of your choosing.
Identify one question that you still have about machine learning that the unit did not answer.

The essay is not a summary of the tutorials. It is your own attempt to organise and interpret what you have learned, and to identify what remains unclear. Honest uncertainty is a sign of good understanding, not a sign of failure.

Mini-project: prepare a short presentation

Mini-project

Prepare a ten-minute presentation that introduces machine learning to a non-technical audience. Your audience could be a group of high school students, business colleagues, or members of the public. Choose the audience that you think would be most useful for you to practise communicating with.

Your presentation should:

Explain what machine learning is, without using jargon.
Give at least two concrete examples of machine learning in action.
Explain what machine learning systems cannot do, and why this matters.
Introduce the idea of failure modes in a way that is understandable to non-specialists.
Finish with one thing you think people should be asking about AI that they often are not.

Deliver the presentation to the group (or record it). Get feedback on whether your explanations were clear, and whether your examples were effective. The ability to explain machine learning to non-specialists is a professional skill that is often more valuable than technical depth. This mini-project is your opportunity to develop it.

Submit your slides and a 500-word reflection on what was easiest and hardest about explaining the material.

11 Unit Summary

Unit 1 in one page

Machine learning is the practice of building systems that learn patterns from data rather than being programmed with explicit rules. It reverses the direction of classical programming: rules are the output, not the input.
Artificial intelligence is the broadest term, machine learning is a subset of AI, and deep learning is a subset of ML. The exception is symbolic AI — expert systems and compilers — which is AI but not ML.
The field’s history includes two AI winters, each following a period of unmet promises. The current deep learning era began in 2012 when large datasets, GPU compute, and algorithmic refinements converged.
The shift from symbolic to statistical AI happened through the 1990s, driven by the knowledge acquisition bottleneck and the brittleness of rule-based systems.
Every learning algorithm has four ingredients: features and labels (from the data), parameters (learned), and loss (a measure of error). The training loop is predict, measure, adjust, repeat.
Six capability areas capture most applications: Perception, Prediction, Language, Ranking, Detection, and Generation. Four conditions predict success — data, stable patterns, high volume, and tolerance for error.
Seven failure modes account for most disappointments: overfitting, underfitting, spurious correlation, data leakage, distribution shift, class imbalance, and feedback loops. None is fundamentally an algorithm problem.
The lifecycle has seven stages and is iterative rather than linear. Data work dominates — roughly 60–70% of a project’s time — with modelling and evaluation together taking only 10–20%.

What comes next

Unit 2 shifts from concepts to data. You will learn the difference between structured and unstructured data, how to handle missing values and inconsistencies, how to construct features from raw data, and how to split a dataset properly into training, validation, and test partitions.

The vocabulary you built in Unit 1 — features, labels, parameters, loss, overfitting, data leakage — is assumed from the first page of Unit 2. If you are not comfortable with it, this is the moment to revisit.

12 Further Reading

Each tutorial in Unit 1 has its own further-reading list. The resources below are the ones most worth returning to as you continue through the course.

Mitchell, T. M. (1997). Machine Learning. McGraw-Hill. Chapter 1.
The source of the formal definition used in T-01. Worth reading once, in full, now that you have the context to appreciate it.
Domingos, P. (2012). “A Few Useful Things to Know About Machine Learning.” Communications of the ACM, 55(10), 78–87.
A short, dense article that names many of the practical lessons Unit 1 touched on: the role of data quality, the failure modes, the gap between theory and practice. Highly recommended before Unit 2.
Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). ““Why Should I Trust You?”: Explaining the Predictions of Any Classifier.” Proceedings of KDD 2016.
The paper where the husky/wolf example from T-07 originates. Worth reading to see how the failure was discovered and how the researchers thought about it.
Russell, S., & Norvig, P. (2021). Artificial Intelligence: A Modern Approach (4th ed.). Pearson.
The standard AI textbook. Its introductory chapters cover the history, the paradigms, and the definitions that Unit 1 has surveyed. Best used as a reference throughout the course rather than read cover-to-cover.
Marcus, G. (2020). “The Next Decade in AI: Four Steps Towards Robust and Hybrid AI.” arXiv:2002.06177.
A counterpoint to the statistical learning orthodoxy. Worth reading before Unit 2 to remember that the field is not settled, and that the hybrid perspective from T-03 and T-04 is an active area of debate.

Ready for Unit 2?

If you scored 7 or above on the quiz and feel comfortable with the key terms, you are ready. Open Unit 2: Data Fundamentals to continue. If your score was lower, or if the confidence audit flagged several items, spend an hour revisiting the relevant tutorials before moving on.