1.1 A New Core Skill for Health Sciences

Each generation of public health workers has had to learn new tools. Statistical software such as SAS, SPSS, Stata and R and bibliographic databases such as MEDLINE each became a skill that employers expected and that separated careful users from careless ones, without replacing professional judgement.

This course starts from the premise that artificial intelligence is becoming a skill of the same kind. Artificial intelligence (AI) refers to computer systems that perform tasks that would require intelligence if people did them, such as recognizing images, interpreting text or making predictions. A large language model (LLM) is an AI system trained on very large collections of text to predict the next piece of text in a sequence. Language models are the engines inside general-purpose assistants such as ChatGPT, Claude, Gemini and Microsoft Copilot, which are examples available in 2026.

The course takes a favourable and practical view. People who use these tools well can learn faster, read more evidence, write more efficiently and carry out analyses that would otherwise be out of reach. The same tools also produce fluent text that is wrong, invent references and agree with mistaken users. Skilled use consists of knowing which outcome is likely for a given task and applying the check that catches it. Lewis, the study assistant on every lesson page in this series, is itself an AI system of this kind, and Lesson 2 uses it as a worked example.

Figure 1.1 shows how the ten lessons of the course build from an understanding of the tools to their responsible use in public health practice.

With the premise and plan of the course set out, the next part turns to evidence on how widely AI tools are already used.

1.2 How Widely AI Is Already Used

Two recent sources estimate how widely AI tools are used: a survey of undergraduate students in the United Kingdom and an analysis of messages sent to the consumer version of ChatGPT.

In February 2025, the Higher Education Policy Institute (HEPI), a United Kingdom think tank, published its Student Generative AI Survey 2025 with the education company Kortext (Freeman, 2025). The survey of 1,041 full-time undergraduates, conducted online in December 2024, found that 92 percent used AI tools in some form, up from 66 percent a year earlier, and that 88 percent had used generative AI for assessments. Only 36 percent said their institution had given them training in AI skills.

Chatterji and colleagues (2025) analysed messages sent to the consumer version of ChatGPT in a working paper published by the National Bureau of Economic Research. They estimated that around 10 percent of the world's adult population had adopted the product by July 2025, that the share of messages unrelated to paid work rose from 53 percent to more than 70 percent, and that practical guidance, seeking information and writing accounted for nearly 80 percent of conversations. The paper discloses that two of its authors were employees of OpenAI, the company that makes ChatGPT, which a careful reader notes when weighing its findings.

Both sources measure use, which is different from skill, and neither studied health sciences students or public health workers directly. Together they show that most students will meet these tools whether or not a course teaches them, and that much of the public now asks an assistant for guidance and information, including health information.

The next part turns from how often the tools are used to what controlled experiments show about the speed and quality of the work they help produce.

1.3 What Experiments Show About Speed and Quality

Widespread use matters only if the tools improve work. Two randomized experiments with professional tasks give the clearest early evidence, and both show where the gains stop.

The tabs that follow describe each experiment in turn and then set out what the two studies share.

Public health work contains many tasks like these, such as drafting briefing notes and summarizing reports. The studies suggest that AI can save substantial time on such work, and that its value depends on the user's ability to judge when an output can be trusted. That ability is what this course calls AI literacy.

1.4 AI Literacy in Canadian Health Policy

AI literacy already appears in federal and provincial guidance on the use of AI in health and in government. This part reviews that guidance and then defines the term as this course uses it.

In January 2025, Health Canada published the Pan-Canadian AI for Health (AI4H) Guiding Principles as shared values for federal, provincial and territorial governments. AI literacy is one of its nine principles. The principle asks that decision-makers, health care professionals, researchers and innovators be "equipped to use AI technologies effectively and responsibly." The Government of Canada's Guide on the use of generative artificial intelligence organizes its advice under the acronym FASTER (Fair, Accountable, Secure, Transparent, Educated and Relevant), and its Educated principle asks public servants to "learn about the strengths, limitations and responsible use of the tools; learn how to create effective prompts and to identify potential weaknesses in the outputs." In British Columbia, the Responsible AI Policy Framework for the B.C. Health Sector (Ministry of Health, June 2026) applies to the Ministry, the health authorities and BC Shared Health Services. Lesson 9 examines these documents in detail.

For this course, AI literacy means the knowledge and habits that allow a person to use AI tools effectively, to judge their output and to use them responsibly. The cards below set out its five parts.

With AI literacy defined, the next part introduces the fictional health authority and the AI pilot that supply the examples in this lesson.

1.5 The Cedar Valley AI Pilot

Box 1.1 describes the case used throughout this lesson: a fictional, mostly rural health authority in British Columbia, its Cedar Valley Connector program for older adults who screen as socially isolated or lonely, and the one-year pilot of an enterprise AI assistant by its population health team. The case also introduces Jordan, a Health Sciences co-op student whose work with the assistant supplies examples in later sections.

The team lead's request that Jordan try the assistant on six everyday tasks is the subject of the next part.

1.6 Worked Demonstration: Six Everyday Tasks

A first trial of an assistant on familiar tasks shows what kind of check each type of output needs. Worked Example 1.1 works through the six tasks that Jordan set, giving for each the prompt, an abridged output and a commentary on accuracy and usefulness, and Table 1.1 at its end summarizes the ratings and the check applied to each task.

The next part draws on these checks to describe the habits that distinguish skilled use of the tools.

1.7 From Use to Skilled Use

A skilled user knows which tasks a tool performs reliably, chooses a check that suits each task, protects confidential data and records what was done. Each habit depends on a working understanding of how the tools produce their output, which is why the rest of this lesson and Lesson 2 explain the technology before later lessons apply it.

Box 1.2 applies these habits to four outputs that each need a different check, and the accordion after it gives the answers.

Section 2 begins the technical part of the lesson by defining artificial intelligence, machine learning and deep learning and by explaining how machines learn from data.

2.1 Three Nested Fields

The terms artificial intelligence, machine learning and deep learning are often used as if they meant the same thing. They name fields that sit inside one another, and knowing which one a tool belongs to tells a user a good deal about how it works and how it can fail.

Artificial intelligence is the broadest field. It covers any computer system built to perform tasks that would require intelligence if people did them, whether the system works by searching through possibilities, applying logical rules or learning from examples. The term was introduced in the 1955 proposal for a summer workshop at Dartmouth College, which Section 4 describes.

Machine learning is the part of AI in which a system learns patterns from data instead of following rules that programmers write. Arthur Samuel of IBM used the term in a 1959 paper describing a checkers program that improved by studying recorded games and playing against itself (Samuel, 1959). In public health terms, a machine learning system is given many examples, finds regularities in them and uses those regularities to make predictions about new cases.

Deep learning is the part of machine learning that uses neural networks with many layers. Section 3 explains how these networks work. Since about 2012, deep learning has produced most of the advances in image recognition, speech recognition and language processing. Generative AI refers to deep learning systems that produce new text, images, audio or code, and large language models are the generative systems that work with text. Lesson 2 explains them in detail.

Figure 1.2 draws these fields as regions nested one inside another, from artificial intelligence, the broadest, to large language models, the narrowest.

The outer field includes many systems that use no learning at all. The next part contrasts such rule-based systems with models that learn patterns from data.

2.2 Rules Written by People and Patterns Learned from Data

Before machine learning became dominant, many AI systems were expert systems: programs that applied large sets of if-then rules written down by human experts. MYCIN, developed at Stanford University in the early 1970s as Edward Shortliffe's doctoral work, used roughly 600 rules to identify the bacteria causing severe infections such as bacteremia and meningitis and to recommend antibiotics. In one evaluation, its recommendations received an acceptability rating of 65 percent, compared with ratings of 42.5 to 62.5 percent for five faculty members. MYCIN was never used in practice, partly because a clinician had to type in all of a patient's information on a shared mainframe computer.

Rules written by people remain common in public health. The Cedar Valley Connector program's referral rule (refer anyone who scores 6 or higher on the three-item UCLA Loneliness Scale) is a rule of this kind, as are most screening eligibility criteria. Such rules are transparent and easy to audit, but someone has to know the rule in advance, and many patterns in health data are too complex to write down.

A learned model takes the opposite approach. Gulshan and colleagues (2016) trained a deep neural network on 128,175 retinal photographs that ophthalmologists had graded for diabetic retinopathy, an eye complication of diabetes. Nobody wrote rules describing what retinopathy looks like. The network adjusted its internal parameters until its outputs matched the experts' grades. On one validation set of 9,963 images, at a threshold chosen for high specificity, it detected referable retinopathy with a sensitivity of 90.3 percent and a specificity of 98.1 percent.

Table 1.2 compares the two approaches on where their knowledge comes from, how they improve, how they explain a decision and how they typically fail.

Table 1.2. Rule-based systems and learned models compared by source of knowledge, means of improvement, explanation of decisions and typical failure.

The vocabulary of machine learning

A few terms recur throughout this course and in any paper that uses machine learning. The cards below define them with the retinopathy example.

With these terms in place, the next part sorts machine learning methods into three groups by the kind of data and feedback they learn from.

2.3 Three Ways Machines Learn

Machine learning methods are usually grouped by the kind of data and feedback they learn from. The tabs below describe the three main groups with public health examples.

Each of the three groups sets its parameters from data. The next part explains how practitioners divide the available data so that a model's performance on new cases can be estimated.

2.4 Training, Validation and Test Sets

A model that has seen an example during training can usually predict that example well, so its performance on the training data says little about how it will perform on new cases. Machine learning practice therefore divides the available data into three parts. The training set is used to fit the model's parameters. The validation set is used to compare versions of the model and to choose settings, such as the size of the network or when to stop training. The test set is held back until the end and used once to estimate how the final model will perform on data it has never seen. A common split assigns about 70 percent of records to training and 15 percent to each of the other two sets, although the proportions vary.

A cooking analogy helps. A cook adjusts a recipe while tasting it (training), asks a friend to compare two versions (validation), and then serves the chosen version to guests whose verdict is final (test). If the cook had let the guests taste every trial batch, their verdict would no longer be an independent judgement of the final dish.

Figure 1.3 shows a typical division of the data into training, validation and test sets, followed by external validation on data from a different setting.

Two problems undermine this design. Data leakage occurs when information from the test set reaches the model during training, for example when photographs of the same patient appear in both sets, or when a feature is recorded only after the outcome has occurred. Leakage makes a model look more accurate than it is. The second problem is that a test set drawn from the same source as the training data shares that source's quirks. External validation, in which the model is tested on data from a different site, population or period, shows whether it transfers.

Zech and colleagues (2018) showed why this matters. They trained deep learning models to detect pneumonia on chest radiographs from three hospital systems. In three of five comparisons, the models performed worse on radiographs from hospital systems other than the one they had been developed on. The models could also identify which hospital system a radiograph came from with high accuracy, which suggests that they could exploit differences between hospitals, such as equipment or how common pneumonia was, in place of features of the disease. Lesson 9 returns to the appraisal of prediction models, and HSCI 230 Lesson 8 covers external validity in epidemiological research.

Box 1.3 applies these points to a vendor's claim about the accuracy of a model offered to the Cedar Valley team.

A held-out test set and external validation show how well a model predicts in new data. The next part considers whether a good prediction also explains why an outcome occurs.

2.5 Prediction and Causal Explanation

Machine learning is very good at prediction, and public health often needs explanation. Hernán, Hsu and Healy (2019) proposed a classification of data science tasks that makes the distinction precise. Description summarizes features of data, such as how many people have a condition. Prediction uses some features to forecast others, such as which people are likely to develop a condition. Counterfactual prediction estimates what would happen under an intervention that did not occur, which is the task of causal inference. The authors argue that description and prediction can be done well with good data and algorithms, while counterfactual prediction also requires expert knowledge about the causal structure that produced the data. Bi and colleagues (2019), in their primer for epidemiologists, likewise present machine learning mainly as a set of tools for prediction, classification and clustering.

Table 1.3 applies the three tasks to questions that the Cedar Valley team might ask about the Connector program and shows what each task requires.

Table 1.3. Description, prediction and counterfactual prediction illustrated with questions about the Cedar Valley Connector program, with what each task requires.

A well-known example shows what happens when a prediction is read as an explanation. Caruana and colleagues (2015) described a rule learned from pneumonia data in the 1990s: patients with asthma had a lower risk of dying from pneumonia. The pattern was real in the data, because patients with pneumonia and asthma were usually admitted directly to intensive care, and the extra care lowered their risk. A model that used the rule to decide who could be treated at home would have sent patients with asthma home, where their risk is higher. The model was accurate for the data it learned from and misleading as a guide to action.

The same reasoning applies to every model in this course, including language models. A model learns whatever its training data contain, including patterns produced by care pathways, by who was measured, and by how outcomes were recorded. When those patterns reflect unequal treatment, the model reproduces the inequality. Lesson 9 examines a widely used algorithm that underestimated the needs of Black patients for this reason, and Lesson 10 examines equity more broadly. HSCI 341 Lesson 1 and HSCI 207 Lesson 3 teach the causal reasoning and causal diagrams that counterfactual questions require, and HSCI 230 Lessons 7 to 11 teach the sources of bias.

Box 1.4 shows Jordan checking an assistant's explanation of supervised and unsupervised learning against the primer by Bi and colleagues (2019).

Box 1.5 asks the reader to classify five questions about the Connector program as description, prediction or counterfactual prediction. The accordion after it gives the answers and addresses a common belief about accuracy and explanation.

Section 3 looks inside deep learning and shows how a neural network learns its parameters from examples.

3.1 The Artificial Neuron

A neural network is a machine learning model built from many simple units, called artificial neurons, connected in layers. The idea goes back to McCulloch and Pitts (1943), who described a simplified mathematical model of a nerve cell that either fires or stays silent depending on its inputs. Modern artificial neurons are loosely inspired by biology, but they are best understood as small mathematical functions.

An artificial neuron does three things. It multiplies each input by a weight, a number that sets how strongly that input counts and in which direction. It adds the weighted inputs together with a bias, a number that shifts the result up or down. It then passes the sum through an activation function, which converts the sum into the neuron's output. The weights and bias are the neuron's parameters, and learning consists of adjusting them.

Figure 1.4 draws a single neuron with two inputs, using illustrative numbers.

Worked Example 1.2 follows the three steps for a neuron that estimates the probability that an older adult will screen as lonely.

Readers who have studied logistic regression will recognize this calculation: a single neuron with a sigmoid activation computes the same function as a logistic regression model. Neural networks become more powerful when many neurons are combined. Several activation functions are in common use. The sigmoid squeezes values into the range 0 to 1 and suits outputs that represent probabilities. The hyperbolic tangent (tanh) squeezes values into the range −1 to 1. The rectified linear unit (ReLU) passes positive values unchanged and sets negative values to zero, and its simplicity makes it the usual choice inside large modern networks. Each of these functions is non-linear, and that non-linearity is what allows a network of neurons to represent curved and irregular patterns.

The next part shows how neurons are combined in layers to form a network.

3.2 From Neurons to Networks

A neural network arranges neurons in layers. The input layer holds the features, such as the pixel values of an image. One or more hidden layers each take the outputs of the previous layer as their inputs, and the output layer produces the prediction. The hidden layers are called hidden because their values are neither given in the data nor reported as the answer; the network invents them as intermediate features.

Figure 1.5 shows a small network of this kind, with two inputs, one hidden layer of four neurons and one output.

Hidden layers matter because a single neuron can only separate two classes with a straight line (or, with more inputs, a flat plane). Many real patterns are not separable in this way. Minsky and Papert showed in 1969 that a single-layer network cannot learn the exclusive-or pattern, in which a point belongs to one class when exactly one of two conditions holds. A network with a hidden layer can combine several straight boundaries into a curved or enclosed one, and with enough hidden neurons it can approximate very complicated boundaries.

The number of parameters grows quickly with the size of the network. The network in the trainer in Interactive 1.1, with two inputs, eight hidden neurons and one output, has 2 × 8 weights and 8 biases in the hidden layer plus 8 weights and 1 bias in the output, for 33 parameters. AlexNet, the image network that Section 4 describes, had about 60 million parameters, and the large language models discussed in Lesson 2 have far more. A network with many hidden layers is called a deep network, which is the origin of the term deep learning.

A network must learn the values of all these parameters from examples, and the next part explains how it does so.

3.3 How a Network Learns

A network starts with random weights, so its first predictions are poor. Learning consists of measuring how wrong the predictions are and adjusting every weight a little in the direction that reduces the error.

The error (also called the loss) is a single number that summarizes how far the network's predictions are from the labels across the training examples. For classification, a common choice is the cross-entropy error, which is large when the network gives a low probability to the correct class and close to zero when it gives a high probability to the correct class. A network that assigns a probability of 50 percent to every example has a cross-entropy error of about 0.69.

Gradient descent reduces the error step by step. Imagine standing on a hillside in thick fog and trying to reach the valley floor. You cannot see the valley, but you can feel which way the ground slopes under your feet, so you take a step downhill and feel again. The gradient is the slope of the error with respect to each weight: it says how much the error would change if that weight were nudged up. Each step moves every weight a small distance against its gradient. The size of the step is set by the learning rate. A rate that is too small makes learning slow, and a rate that is too large makes the network overshoot the valley and bounce around it.

Box 1.6 states the rule that each step of gradient descent applies to every weight.

The difficulty in a network with hidden layers is working out the slope for every weight, including weights deep inside the network whose effect on the error passes through many other neurons. Backpropagation solves this problem. It computes the error at the output and then works backward through the network, layer by layer, using the chain rule of calculus to assign each weight its share of responsibility for the error. Rumelhart, Hinton and Williams described the procedure and showed what it could achieve in a paper published in Nature on 9 October 1986. Related mathematical methods had been described earlier, but their paper showed that backpropagation allowed networks with hidden layers to learn useful internal features, and it became the standard way of training neural networks.

Worked Example 1.3 applies one step of the update rule in Box 1.6 to the neuron in Worked Example 1.2, with a learning rate of 0.1.

Training repeats this process many times. One pass through all the training examples is called an epoch. Large networks are usually trained on small random batches of examples at a time, and a model may see each example many times before its error stops falling.

Box 1.7 shows Jordan asking the pilot assistant to work through the same step and checking each number before relying on it.

The next part applies gradient descent to a small network that the reader can train in the browser.

3.4 Train a Network in Your Browser

Interactive 1.1 builds a network with two inputs (the horizontal and vertical position of a point in a square), one hidden layer of tanh neurons and a sigmoid output. It learns to separate teal circles (class 1) from red squares (class 0) in one of three patterns; the Four quadrants pattern is a version of exclusive-or. Each training pass applies one step of gradient descent using all 100 training points. A separate test set of 200 points, drawn from the same pattern, is never used for training and measures how well the network generalizes. The left panel shades each region of the square by the network's prediction, and the right panel plots the training error (solid teal) and test error (dashed red) as training proceeds.

Interactive 1.1. A trainer for a small neural network that learns to separate teal circles from red squares in one of three patterns and plots the training and test error as it learns.

Box 1.8 sets out four experiments with the trainer, and the accordion after it explains what each experiment shows.

The experiments show, in a network small enough to watch, what happens when a model is too simple, when its learning rate is poorly chosen and when it fits the noise in its training data. The next part explains these results in terms of overfitting and generalization.

3.5 Overfitting and Generalization

The purpose of training is generalization: good performance on new data that the model has not seen. Overfitting occurs when a model learns the details and noise of its training data so closely that its performance on new data gets worse. Its signature is the pattern in Experiment 4, where training error keeps falling while validation or test error rises. Underfitting is the opposite problem: a model too simple to capture the pattern, as in Experiment 2, performs poorly on both training and new data.

Overfitting is most likely when a model has many parameters relative to the number of training examples and when the labels contain errors. Practitioners reduce it in several ways. They collect more training data, use a simpler model, stop training when the validation error begins to rise (early stopping), add a penalty that keeps weights small (regularization), or randomly switch off a share of neurons during each training step (dropout) so that the network cannot rely on any single path. Each of these choices is made with the validation set, and the test set is kept for the final estimate.

Generalization has a second meaning that matters more in public health. A model can generalize well to a test set drawn from the same hospital and year as its training data and still fail elsewhere. The pneumonia models described in Section 2 performed worse at hospital systems other than their own, and a model trained on urban patients may perform poorly in a rural health authority such as Cedar Valley. A change in the population, equipment, coding practices or behaviour that produced the data, often called distribution shift, can degrade a model that once performed well. Lesson 9 shows how this happened to a well-known surveillance model.

Because overfitting is most likely when a model has many parameters relative to the number of training examples, the next part turns to why deep networks need large amounts of data and computing power.

3.6 Why Deep Networks Need Data and Computing Power

A deep network has millions or billions of parameters, and each one must be set from data. With too few examples, a network of that size can simply memorize them, which is the overfitting problem on a large scale. Deep learning therefore advanced only when large labelled datasets became available. The retinopathy network in Section 2 was trained on 128,175 photographs, each graded by ophthalmologists, and the ImageNet image collection described in Section 4 provided about 1.2 million labelled training images for its 2012 competition. Labelling is often the most expensive part of a health project, because it requires clinical experts.

Training also requires a very large number of arithmetic operations, since every training step updates every parameter. Graphics processing units (GPUs), designed for video games, perform thousands of simple multiplications at once, and their use made deep networks practical; AlexNet was trained on two of them. The language models in Lesson 2 are trained on far larger collections of text, using thousands of specialized processors, at a cost in money and energy that only large organizations can bear.

Two practices make deep learning usable in settings with less data. In pretraining and fine-tuning, a network is first trained on a very large general dataset and then trained further on a small dataset for a specific task, so that it starts with useful features already learned. The general-purpose assistants in this course are pretrained models of this kind. In other settings, simpler models are the better choice. For structured clinical data, such as tables of patient characteristics, a systematic review found no performance benefit of machine learning methods over logistic regression for clinical prediction models (Christodoulou and colleagues, 2019). Deep learning has its clearest advantages with images, sound and text, where the useful features are hard to specify in advance.

The accordion that follows addresses three common beliefs about neural networks: that they work the way the brain does, that the network that fits its training data best is the best network, and that deep learning always beats simpler methods.

Section 4 traces how these ideas developed, from the first artificial neurons and the perceptron through decades of disappointment to the deep learning and generative models of today.

4.1 Foundations, 1943 to 1960

The history of AI is short enough to fit within a single lifetime, and its early ideas still shape the tools students use today. Section 3 described the first of them: the mathematical neuron of McCulloch and Pitts (1943).

In 1950, the British mathematician Alan Turing published "Computing Machinery and Intelligence" in the journal Mind. Turing replaced the question "Can machines think?", which he considered too vague to answer, with a test he called the imitation game, now known as the Turing test. In the game, a human judge holds written conversations with a person and a machine and tries to tell which is which. Turing also suggested that it might be easier to build a machine that learns, in the way a child does, than to program adult intelligence directly, an idea that anticipated machine learning.

The field took its name from a proposal dated 2 September 1955, in which John McCarthy of Dartmouth College, Marvin Minsky, Nathaniel Rochester of IBM and Claude Shannon of Bell Laboratories asked for funding for a summer research project on "artificial intelligence". The proposal rested on the conjecture that every aspect of learning or intelligence could in principle be described so precisely that a machine could simulate it. The Dartmouth workshop took place over several weeks in the summer of 1956 and is usually treated as the founding event of AI as a research field.

Two years later, Frank Rosenblatt, a psychologist at the Cornell Aeronautical Laboratory, described the perceptron in Psychological Review (Rosenblatt, 1958). A perceptron is a single layer of artificial neurons whose weights are adjusted automatically when it makes a mistake, which makes it one of the first neural networks that learned from examples. Rosenblatt first simulated it on an IBM 704 computer in 1957. His team then built a machine, the Mark I Perceptron, which used a grid of 400 light sensors as its "retina" and was funded by the United States Office of Naval Research; it was first demonstrated publicly in 1960. A 1958 press event led to newspaper reports that the Navy expected such machines eventually to walk, talk, see and reproduce themselves. The claims far exceeded what a single-layer network could do, and the gap between them and the results would matter in the following decade.

4.2 The AI Winters

The early decades of AI combined real progress with predictions that went unmet, and funders eventually responded. The periods of reduced funding and interest that followed are known as AI winters. Historians date them differently, but most identify two.

The accordion that follows describes each winter and the events that led to it.

The AI winters are worth knowing for two reasons. They show that claims for a new technology can run well ahead of its results, which is a reason to judge any tool by evidence on the specific task at hand. They also show that ideas set aside during a winter can return when conditions change. Neural networks, out of favour for long stretches between 1969 and the late 2000s, became the basis of modern AI once large datasets and fast hardware were available.

Figure 1.6 places selected events on a timeline, with the two AI winters shaded and the deep learning and generative era marked in red.

The return of neural networks depended on large datasets and fast hardware. The next part describes how those conditions came together in the deep learning breakthrough of 2012.

4.3 Data, Graphics Processors and the Deep Learning Breakthrough

By the late 2000s, two conditions that early researchers had lacked were in place: large labelled datasets and affordable hardware for fast parallel arithmetic. ImageNet, a collection of millions of photographs labelled by the objects they contain, supported an annual competition, the ImageNet Large Scale Visual Recognition Challenge, in which teams' programs labelled images they had not seen.

In 2012, a team from the University of Toronto (Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton) entered a deep convolutional neural network, later called AlexNet. Its top-5 error rate, the proportion of test images for which the correct label was not among the network's five best guesses, was 15.3 percent, compared with 26.2 percent for the second-best entry. The network had about 60 million parameters, was trained on about 1.2 million labelled images, and was trained on two graphics processors. The size of the margin persuaded much of the field that deep networks trained on large datasets could outperform methods built on hand-designed features, and deep learning spread quickly to speech recognition, translation and medical imaging, including the retinopathy study described in Section 2.

Reinforcement learning made a similar public leap in March 2016, when AlphaGo defeated the Go champion Lee Sedol four games to one in Seoul. Go had long been considered too complex for computers to play at the highest level.

These successes came from systems that classify, predict and play games, and the next part turns to systems that generate new content.

4.4 The Generative Turn

The generative turn produced the general-purpose assistants that this course examines. This part follows it from a new network architecture for text to the public release of ChatGPT, and then notes what deep learning has done for biology.

In June 2017, Vaswani and colleagues introduced the transformer, a network architecture built around a mechanism called attention, which lets the model weigh the relevance of every part of a text to every other part. Transformers could be trained efficiently on graphics processors, which made it practical to train language models on very large collections of text. Lesson 2 explains tokens, attention and the training that turns such a model into an assistant.

On 30 November 2022, OpenAI released ChatGPT to the public. It was reported to have gained one million users in five days and 100 million within about two months, and Section 1 described its later reach. Assistants from several other companies followed, along with models that work through problems in steps before answering and agents that use tools to complete tasks, which Lesson 2 describes. This shift from systems that classify and predict to systems that generate text, images and code is often called the generative turn, and it is the reason AI literacy has become a concern for every student and professional.

Deep learning has also changed biology in ways that matter for health. Protein structure prediction, which estimates the three-dimensional shape of a protein from its sequence of amino acids, had resisted decades of effort. Deep learning systems, most prominently AlphaFold from Google DeepMind, made such predictions far more accurate, and computational methods now also design new proteins. Protein structures guide research on medicines and vaccines, which is why this work received the 2024 Nobel Prize in Chemistry.

The next part describes the contribution of Canadian researchers and institutions to the developments traced in this section.

4.5 Canada's Part in the History

Canadian researchers and institutions played a large part in the deep learning era. Geoffrey Hinton, a professor at the University of Toronto, was a co-author of the 1986 backpropagation paper, co-developed the Boltzmann machine and supervised the AlexNet team. Yoshua Bengio of the Université de Montréal and Mila made foundational contributions to deep learning and to neural models of language. Richard Sutton of the University of Alberta and Amii is a founder of modern reinforcement learning. The Canadian Institute for Advanced Research (CIFAR), which hosts the Learning in Machines & Brains research program, has long supported this community.

In 2017, the federal government launched the Pan-Canadian Artificial Intelligence Strategy, described by CIFAR as the first national AI strategy in the world. Budget 2017 committed $125 million, and CIFAR led the strategy's first phase, which aimed to attract, retain and train AI researchers. The strategy is built around three national AI institutes: Amii in Edmonton, Mila in Montreal and the Vector Institute in Toronto. A second phase, announced in Budget 2021 with $443.8 million, added support for the adoption and use of AI.

The flip cards that follow describe four recent awards for work this section has traced: the Turing Awards of 2018 and 2024 and the 2024 Nobel Prizes in Physics and Chemistry.

For students in British Columbia, this history is close to home. Several of the methods that power the tools in this course were developed at Canadian universities, and the national institutes continue to train researchers in the field. The same period has seen Canadian governments adopt the guidance described in Section 1, including the Pan-Canadian AI for Health Guiding Principles and the B.C. framework for the health sector.

The final part of this section considers what this history suggests for people who use AI tools.

4.6 What the History Suggests for Users

Three lessons from this history carry into the rest of the course. Progress in AI has depended as much on data and computing power as on new ideas, which is why the capabilities of a tool reflect what it was trained on. Claims for new systems have often run ahead of their results, which is why each lesson asks what the evidence shows for a specific task. Change has accelerated: the time from the transformer paper to a widely used public assistant was about five years, and the tools have continued to change since. Skilled users therefore treat their knowledge of particular products as provisional and keep learning, which Lesson 10 makes into a plan.

Box 1.9 asks the reader to check a five-line history of AI written by an assistant against this section, and the accordion after it gives the answer.

Box 1.10 draws the lesson together, and Table 1.4 within it sets out how machine learning systems and general-purpose assistants perform on six kinds of task and the check that applies to each.

A good next step is to return to the trainer in Section 3 and try settings that the experiments did not cover, such as the Four quadrants pattern with two, four and eight hidden units, and to ask an assistant to explain what you observe and then check its explanation against this lesson. Lesson 2 shows how the same ideas, applied to text at very large scale, produce the language models inside today's assistants.

Bringing It All Together

This lesson opened the course with its premise: the ability to use AI and large language models well is becoming a core skill in health sciences and public health, much as statistical software and bibliographic databases became core skills in earlier decades. Use is already widespread, controlled experiments show large gains in speed and quality on suitable tasks, and Canadian health policy now names AI literacy as a principle. The same experiments, and Jordan's test of six everyday tasks in the fictional Cedar Valley pilot, show that the gains depend on knowing where a tool is reliable and on applying a check suited to each task.

The lesson then explained the technology on which every later lesson builds. Machine learning systems learn patterns from data instead of following written rules, through supervised, unsupervised or reinforcement learning, and their performance must be judged on data they have not seen, ideally from a different setting. Neural networks learn by adjusting their weights through gradient descent and backpropagation, and the trainer showed how they can underfit, learn well or overfit. Deep networks need large datasets and fast hardware, which is why deep learning advanced only after 2012.

The history in Section 4 placed these ideas in time, from Turing and Dartmouth through the AI winters to ImageNet, the transformer and ChatGPT, and showed Canada's large part in the deep learning era. The history and the technical sections point to the same practical conclusion: these systems are powerful pattern learners whose output reflects their training data, so skilled users take advantage of their strengths and check their output.

The final reflection asks you to explain how machine learning works, why models may not transfer between settings, and how a team should use an AI assistant, for an audience of new colleagues.

Final Knowledge Assessment

This glossary defines the terms, frameworks and people introduced in this lesson; later lessons in the course build on them.