L1
·
Quiz
·
Lab
L2
·
Quiz
·
Lab
L3
·
Quiz
·
Lab
L4
·
Quiz
·
Lab
Module Test
Lesson 1 · Module 3

Training Data — The Raw Material

Before an AI can find patterns, someone has to gather an ocean of examples.
Where does all the data come from — and why does its quality matter so much?

In 2009, Google engineers announced they could predict flu outbreaks across the United States by analyzing which search queries people typed — terms like "fever," "aching muscles," and "flu medication." Their model tracked 45 search term combinations and matched CDC flu reports with startling accuracy. The media called it a triumph of big-data pattern matching. Then, in 2013, the system dramatically overestimated flu prevalence by a factor of two. Investigators found that Google had quietly changed its search algorithm, altering which suggestions appeared as users typed — changing the very data stream the model depended on. The patterns it had learned were tied to a platform's behavior, not just human biology.

What Is Training Data?

Every AI model that learns from experience — rather than following hand-coded rules — begins with a dataset. A dataset is a structured collection of examples: images paired with labels, sentences paired with translations, medical scans paired with diagnoses, user clicks paired with purchase outcomes.

The model studies these examples statistically, extracting recurring patterns. It never truly "understands" the examples the way a person might; it finds mathematical regularities. So the dataset is everything — it is the entire reality the AI gets to see during training. Whatever biases, gaps, or errors exist in that data will shape the model's behavior long after training ends.

15T
tokens in GPT-4's training data (est.)
5.8B
images in LAION-5B dataset
3.5B
labeled images in ImageNet lineage

Where Does Training Data Come From?

Modern AI datasets draw from several major sources:

The open web. Common Crawl is a nonprofit that has archived petabytes of web pages since 2008. GPT-2, GPT-3, and most large language models draw heavily from it. One major sub-corpus, WebText, was built entirely from Reddit posts that received at least three upvotes — effectively letting Reddit's voting system act as a quality filter.

Books. The Books1 and Books2 corpora used in GPT-3 training contained hundreds of thousands of digitized novels and nonfiction texts, giving language models exposure to long-form coherent prose.

Wikipedia. At roughly 20 billion words in English alone, Wikipedia is clean, structured, and edited — making it a high-quality anchor in many training mixes.

Human-labeled data. For tasks requiring precision — medical imaging, legal document review, handwriting recognition — teams of human annotators label examples manually. Amazon Mechanical Turk and similar platforms have employed millions of workers globally to create labeled training sets.

Case: ImageNet (2009)

Princeton researcher Fei-Fei Li and her team used Amazon Mechanical Turk workers in 167 countries to label 14 million images across 21,841 categories. The result, ImageNet, became the benchmark that triggered the deep learning revolution when Alex Krizhevsky's AlexNet model slashed the ImageNet error rate from 26% to 15.3% in 2012. The data collection effort itself took over two years.

Why Data Quality Defines AI Behavior

The Google Flu Trends failure illustrated a principle engineers call distribution shift — when the data the model encounters after deployment differs from the data it trained on. The training patterns were real but fragile: they depended on a stable relationship between search queries and flu behavior that changed when the platform changed.

A more systemic form of this problem is historical bias. In 2018, researchers at MIT and Microsoft published a study showing that commercial facial recognition systems from IBM, Microsoft, and Face++ had error rates up to 34.7 percentage points higher for dark-skinned women than for light-skinned men. The models had been trained on datasets of faces that skewed heavily white and male — and they learned patterns that reflected that skew.

Key Principle

An AI model cannot be fairer, more accurate, or more representative than the data it was trained on. "Garbage in, garbage out" is the oldest rule in computing — in machine learning, it applies at civilizational scale.

Training dataThe collection of labeled or unlabeled examples a model learns from before deployment.
Distribution shiftWhen the real-world data a model encounters differs significantly from the data it was trained on, causing errors.
AnnotationThe process of human workers labeling raw data — adding tags, categories, or judgments — so a model can learn from it.

Lesson 1 Quiz

Training Data — The Raw Material
1. Why did Google Flu Trends dramatically overestimate flu cases in 2013?
Correct! The platform's autocomplete behavior changed, altering which terms users typed — a classic case of distribution shift breaking a trained model.
Not quite. The failure was a distribution shift: Google's platform changes altered the data stream the model's patterns depended on.
2. What was the main quality-control mechanism used to build WebText (a GPT-3 training corpus)?
Exactly right. Reddit's crowd-voting acted as a quality proxy, though it also introduced Reddit's own demographic skews into the training data.
Actually, OpenAI used Reddit upvotes as the filter — posts with at least 3 upvotes, meaning the crowd's voting judgment decided what was "quality."
3. The MIT/Microsoft 2018 study on facial recognition found the worst performance for which group?
Correct. Error rates were up to 34.7 percentage points higher for dark-skinned women than light-skinned men — reflecting training datasets skewed toward lighter-skinned, male faces.
The research by Joy Buolamwini and Timnit Gebru found dark-skinned women had the highest error rates — up to 34.7 percentage points worse than light-skinned men.
4. What does "distribution shift" mean in machine learning?
Exactly. Distribution shift is one of the most common causes of AI failure in production — the world moves on, but the model's learned patterns stay frozen in time.
Distribution shift refers to the mismatch between training data and real-world deployment data — when the world changes and the model's learned patterns no longer fit.

Lab 1 · Training Data Explorer

Ask the AI assistant about training data sources, biases, and real-world failures.

Your Mission

You've learned that training data shapes everything. Now interrogate those ideas. Ask about specific datasets, data collection methods, bias origins, or the Google Flu Trends / ImageNet cases. Dig into anything from this lesson.

Try asking: "Why did Reddit upvotes make WebText biased?" — or — "How could ImageNet have been collected more fairly?"
Training Data Lab
L1 · Pattern Party
Lesson 2 · Module 3

Supervised Learning — Teaching by Example

The most common way AI learns is also the most human: show it thousands of correct answers and let it figure out the rules.
How does repeatedly correcting an AI's mistakes eventually make it accurate?

In October 2012, a doctoral student named Alex Krizhevsky submitted an entry to ImageNet's annual Large Scale Visual Recognition Challenge. His neural network, later called AlexNet, had been trained for about a week on two NVIDIA GTX 580 graphics cards — consumer gaming hardware — using 1.2 million labeled images. AlexNet achieved a top-5 error rate of 15.3%, crushing the second-place entry at 26.2%. Every image it had trained on came with a human-supplied label. Every time it guessed wrong during training, an algorithm called backpropagation adjusted its internal weights to make a better guess next time. It did this roughly 90 epochs — 90 complete passes through the entire dataset — before it was ready.

The Basic Idea

Supervised learning is training with labeled examples. Each training example consists of an input — an image, a sentence, a set of sensor readings — paired with the correct output a human has assigned. The model makes a prediction, compares it to the correct answer, measures its error, and adjusts its internal parameters to reduce that error. Repeat millions or billions of times.

The algorithm that performs the adjustment is typically called gradient descent: mathematically walking downhill on an error landscape, step by step, toward lower and lower mistake rates. The process that calculates which direction is "downhill" through all the model's layers is called backpropagation, invented in its modern form by David Rumelhart, Geoffrey Hinton, and Ronald Williams in 1986.

Classification Tasks

The model learns to assign inputs to categories. Is this email spam or not-spam? Does this X-ray show a tumor? Which digit is this handwritten number? The training label is a category name.

Regression Tasks

The model learns to predict a continuous value. What will this house sell for? How many units will we ship next quarter? The label is a number, and the model minimizes the gap between its guess and the true value.

A Real Journey: Spam Detection

One of the earliest and most consequential supervised learning deployments was email spam filtering. In 2002, Paul Graham published "A Plan for Spam," describing a Bayesian probabilistic classifier trained on personal email — legitimate messages labeled "ham," junk labeled "spam." The model learned that certain word combinations had high spam probability. Users who trained the filter on their own mail got personalized accuracy.

By 2004, Microsoft Research's teams and Google's Gmail engineers were training neural classifiers on hundreds of millions of labeled messages. The "label" was implicit: messages users marked as spam or moved to inbox. This is sometimes called implicit supervision — users' actions become labels without anyone explicitly sitting down to annotate data.

Case: MNIST — The "Hello World" of Machine Learning

Created in 1998 by Yann LeCun, Corinna Cortes, and Christopher Burges at Bell Labs, MNIST is a dataset of 70,000 handwritten digits (0–9) with human-supplied labels. It became the standard benchmark for testing new supervised learning algorithms. As of 2024, the best models achieve error rates below 0.2% on MNIST — essentially matching human performance on an extremely clean task. The dataset's clarity made it useful precisely because the labels were unambiguous: a "7" is a 7.

The Limits of Supervision

Supervised learning has a fundamental cost: someone has to label everything. For ImageNet's 14 million images, that took two years of crowdsourced labor. For medical AI systems, labels must come from certified specialists — making them expensive. For rare diseases or unusual events, labeled examples may simply not exist in sufficient quantity.

When a model is trained to a very high accuracy on training data but fails on new examples, we call this overfitting — the model has memorized the specific examples rather than learning the general pattern. Preventing overfitting while maximizing accuracy on unseen data is the central challenge of supervised learning engineering.

The Overfitting Analogy

Imagine studying for an exam by memorizing the exact questions from last year's test — verbatim. You'd score 100% on that old exam, but fail completely if the teacher changed any question. Overfitting is the mathematical equivalent: perfect memory, poor generalization.

Supervised learningTraining a model on input-output pairs where correct outputs (labels) are provided by humans.
BackpropagationThe algorithm that calculates how to adjust a neural network's weights to reduce prediction error, flowing corrections backward through layers.
OverfittingWhen a model learns training examples so precisely that it fails to generalize to new, unseen data.

Lesson 2 Quiz

Supervised Learning — Teaching by Example
1. What hardware did Alex Krizhevsky use to train AlexNet, the model that won ImageNet 2012?
Correct! Krizhevsky used consumer-grade gaming GPUs — the same cards people used to play video games — to achieve a breakthrough that reshaped the entire field.
AlexNet was trained on two NVIDIA GTX 580 graphics cards — gaming hardware at the time. The result demonstrated GPUs could massively accelerate neural network training.
2. In supervised learning, what happens each time a model makes a wrong prediction during training?
Exactly right. Backpropagation calculates the direction to adjust each weight, and gradient descent steps the model toward lower error — one correction at a time, millions of times.
Each wrong prediction triggers backpropagation — the algorithm calculates how to adjust internal weights so the model does better next time. No human intervention needed during training.
3. Paul Graham's 2002 spam filter used "implicit supervision" from users. What served as the training labels?
Correct! When users clicked "Mark as Spam," that action became a label. Millions of such user actions collectively trained the classifier — implicit supervision at scale.
The labels were implicit — user behaviors like clicking "spam" or moving messages created the training signal without anyone explicitly sitting down to annotate a dataset.
4. A student memorizes every exact question from last year's exam but fails when new questions are asked. This is analogous to which machine learning problem?
Exactly. Overfitting is learning the training examples too precisely — achieving high accuracy on the data seen during training while failing to generalize to new inputs.
This describes overfitting — when a model memorizes training examples rather than learning generalizable patterns, causing it to fail on new data.

Lab 2 · Supervised Learning Sandbox

Explore backpropagation, overfitting, and how labeled examples teach AI models.

Your Mission

Supervised learning is the backbone of most deployed AI. Ask about how backpropagation works, why overfitting happens, how many training epochs AlexNet needed, or how spam filters use implicit labels. Challenge the AI to explain gradient descent in plain language.

Try asking: "Explain gradient descent like I'm 12 years old." — or — "How would you prevent overfitting in a medical imaging classifier?"
Supervised Learning Lab
L2 · Pattern Party
Lesson 3 · Module 3

Beyond Labels — Unsupervised & Reinforcement Learning

What happens when there are no right answers to copy from? AI can still discover structure — or learn by trial, error, and reward.
Can a machine learn without being told what's correct?

In October 2015, DeepMind's AlphaGo defeated the European Go champion Fan Hui five games to zero — the first time a computer had beaten a professional Go player on a full 19×19 board without handicap. AlphaGo was not given a list of rules for good Go strategy. Instead, it was trained first on a dataset of 30 million human expert moves — supervised learning — and then underwent reinforcement learning by playing millions of games against copies of itself, receiving a positive signal for winning and a negative signal for losing. The Go board's near-infinite configuration space made rule-based programming impossible. The algorithm had to discover strategy on its own through consequence.

Unsupervised Learning — Finding Structure Without Labels

In unsupervised learning, the model receives data with no labels at all. Its task is to find structure, clusters, or patterns on its own. There is no correct answer to compare against; the model must discover organization the data already contains.

Clustering algorithms, like k-means, group data points that are mathematically similar. Netflix's earliest recommendation work used collaborative filtering — clustering users by viewing behavior without anyone labeling user "types."

Dimensionality reduction techniques like PCA (Principal Component Analysis) and t-SNE compress high-dimensional data into lower-dimensional representations, revealing structure invisible in raw numbers. Biologists use t-SNE to visualize which gene expression patterns cluster together across thousands of cells.

Clustering (e.g., K-Means)

Groups similar data points together. Used in customer segmentation, document topic modeling, and genetic analysis — anywhere you want to discover natural groupings humans haven't pre-defined.

Autoencoders

Neural networks trained to compress data into a small representation and then reconstruct it. The bottleneck forces the model to learn what matters most. Used for anomaly detection: an input that reconstructs poorly is probably unusual.

Reinforcement Learning — Learning by Consequence

Reinforcement learning (RL) takes a different approach entirely. An agent operates in an environment, takes actions, and receives rewards or penalties. Through trial and error across many episodes, it learns a policy — a strategy for choosing actions that maximize cumulative reward over time.

Unlike supervised learning, there is no teacher providing correct answers. Unlike unsupervised learning, there is feedback — but it comes from the environment's rules, not human labels on data.

Case: OpenAI's Dota 2 Bot (2019)

OpenAI Five played 45,000 years of Dota 2 against itself in a single month using 256 GPUs and 128,000 CPU cores. The reward signal was simply winning or losing. No human ever labeled a good move or a bad move. By 2019, OpenAI Five defeated the reigning world champion team OG two games to zero. The agents discovered team coordination, resource management, and sacrifice strategies entirely through accumulated consequence — millions of games of trial, error, and reward.

Where These Methods Meet: Modern Large Models

Today's most powerful AI systems blend all three paradigms. GPT-4 was first pretrained on web text using a form of self-supervised learning (predicting the next word — which is unsupervised in spirit). It was then fine-tuned with supervised learning on human-written examples. Finally, it was shaped by Reinforcement Learning from Human Feedback (RLHF) — human raters scored outputs, and those scores became a reward signal for further training.

RLHF was introduced by OpenAI researchers in a 2017 paper and became central to making ChatGPT helpful rather than merely technically accurate. The human raters essentially acted as the "environment" giving rewards — making real people part of the reinforcement learning loop.

1959
Arthur Samuel coins "machine learning" while building a checkers-playing program that improves through self-play — an early RL ancestor.
1992
TD-Gammon by Gerald Tesauro learns backgammon at expert level through reinforcement learning — no human strategy encoded.
2013
DeepMind's DQN learns 49 Atari games from raw pixel input and reward score alone — superhuman on 29 of them.
2017
AlphaGo Zero learns Go from scratch with zero human data — only the rules and self-play — surpassing AlphaGo in 40 days.
2022
ChatGPT launches with RLHF as a key training stage — human preferences become the reward signal.
Unsupervised learningTraining on unlabeled data to discover hidden structure, clusters, or compressed representations without human-provided correct answers.
Reinforcement learningAn agent learns a policy by taking actions in an environment and receiving rewards or penalties based on outcomes.
RLHFReinforcement Learning from Human Feedback — using human preference ratings as the reward signal to shape a model's behavior.

Lesson 3 Quiz

Unsupervised & Reinforcement Learning
1. How did AlphaGo learn Go strategy — what were its two main training stages?
Correct! AlphaGo first learned from human expert moves (supervised), then refined its strategy by playing millions of games against itself (RL) — winning only was the reward.
AlphaGo used supervised learning on 30 million human expert moves first, then reinforcement learning through self-play where winning was the reward signal.
2. What was the reward signal in OpenAI Five's Dota 2 training?
Exactly. The only signal was the final outcome — win or loss. The agents discovered strategy, coordination, and sacrifice through billions of episodes of trial and error.
OpenAI Five used only win/loss as its reward — no move-by-move human labeling. Strategy emerged entirely from playing 45,000 simulated years of games.
3. What is an autoencoder used for in unsupervised learning?
Right! The bottleneck forces the model to learn what matters. An input that reconstructs poorly — high reconstruction error — is likely anomalous, which makes autoencoders valuable for fraud and failure detection.
Autoencoders compress data through a bottleneck and reconstruct it. Inputs that reconstruct poorly are probably anomalous — making them useful for anomaly detection without labeled examples of "anomaly."
4. What role did human raters play in training ChatGPT via RLHF?
Correct. RLHF turns human preferences into a reward signal — raters judge which responses are better, and those judgments train a reward model that guides further RL fine-tuning.
In RLHF, human raters compare model outputs and their preferences train a "reward model." That reward model then guides reinforcement learning — humans become the environment providing rewards.

Lab 3 · RL & Unsupervised Explorer

Dig into reinforcement learning rewards, self-play, RLHF, and unsupervised clustering.

Your Mission

You've learned that AI can discover patterns without labels and learn strategy from nothing but win/loss signals. Ask the AI to explain any of these concepts more deeply — RLHF, how AlphaGo Zero worked differently from AlphaGo, what a policy is in RL, or how clustering reveals hidden groups in data.

Try asking: "How is AlphaGo Zero different from AlphaGo?" — or — "Can you explain what a policy is in reinforcement learning with a real-world analogy?"
RL & Unsupervised Lab
L3 · Pattern Party
Lesson 4 · Module 3

When Patterns Break — AI Failure Modes

Learning from patterns is powerful. But patterns can be spurious, brittle, or dangerously wrong — and real systems have failed in ways worth understanding.
How do we know when to trust what an AI has learned?

In 2016, researchers at the University of Washington published a study showing that a deep learning skin cancer classifier performed impressively on test images — correctly identifying malignant melanomas at dermatologist-level accuracy. Then they examined what the model was actually using to make its decisions. Using a technique called LIME (Local Interpretable Model-Agnostic Explanations), they found that the classifier had learned to associate ruler markings in medical photographs with malignancy. Dermatologists routinely place rulers next to suspicious lesions to document size — meaning malignant-lesion images happened to contain rulers more often than benign-lesion images. The model learned a spurious correlation, not the clinical pattern. It passed the accuracy test for completely the wrong reasons.

Spurious Correlations — Learning the Wrong Pattern

When a model learns something that correlates with the right answer in training data but doesn't causally explain it, the result is a spurious correlation. The model works — sometimes brilliantly — until it encounters data where the spurious feature and the real answer come apart.

In 2020, MIT researchers tested several COVID-19 diagnostic AI models that had been trained on chest X-rays from different hospitals. Many had learned to identify scanner make, image resolution, or patient positioning artifacts that happened to correlate with COVID status in their training sets — because some hospitals had more COVID patients and also happened to use particular imaging equipment. These patterns had nothing to do with the virus and would fail immediately if deployed at a different hospital.

Case: Amazon's Recruiting AI (2018)

Reuters reported in October 2018 that Amazon had quietly abandoned a machine learning recruiting tool after discovering it systematically downgraded resumes from women. The model had been trained on ten years of Amazon's own hiring decisions — predominantly male hires, reflecting the tech industry at the time. The model learned that male-associated terms and features correlated with successful hiring, so it penalized resumes that included the word "women's" (as in "women's chess club") and downgraded graduates of two all-women colleges. Amazon scrapped the project rather than deploy a system it could not make gender-neutral.

Adversarial Examples — Patterns Fooled by Tiny Changes

In 2014, Ian Goodfellow, Jonathon Shlens, and Christian Szegedy at Google demonstrated that adding imperceptible noise to an image — changes invisible to human eyes — could cause a neural network to completely change its classification. A panda image with carefully computed pixel noise was classified as a gibbon with 99.3% confidence. The modified image looked identical to humans.

This revealed something profound: neural networks learn patterns that are mathematically real but perceptually alien — they pick up on statistical features in pixel space that don't map to how humans see the world. This isn't a bug in one model; it appears to be a structural property of gradient-descent-trained classifiers.

In 2019, researchers showed that printed stickers placed on a stop sign could cause a self-driving car's classifier to misread it as a speed limit sign at specific angles. Real physical adversarial patches — not just digital noise — could fool deployed systems.

Shortcut Learning

Models exploit the fastest statistical path to reducing training error, not the most meaningful one. A "horse" classifier might learn to detect watermarks from a specific photo archive rather than actual horses — because every horse image happened to come from that archive.

Dataset Contamination

When test data leaks into training data, models appear to perform better than they actually do. GPT models' benchmark scores have been questioned because pre-training on the web may include the benchmark questions themselves.

What Auditing and Interpretability Offer

The skin cancer classifier failure was caught because researchers asked why the model made its predictions — not just whether it was accurate on a test set. This field is called explainable AI (XAI) or interpretability.

Techniques like LIME and SHAP (SHapley Additive exPlanations) try to attribute a model's decision to specific input features. Attention visualization shows which words a language model attends to most when generating an answer. None of these techniques is perfect — all are approximations — but they have caught real errors before deployment.

In 2022, the EU's AI Act proposed mandatory conformity assessments for "high-risk" AI systems — medical, hiring, law enforcement — requiring documentation of training data, testing procedures, and bias audits before deployment. The ruler-in-the-photo problem is precisely what such audits are designed to catch.

The Core Lesson

Accuracy on a test set is necessary but not sufficient. A model can score 95% and be learning completely the wrong thing. Understanding what patterns an AI has actually learned — not just how often it's right — is the frontier challenge of machine learning safety.

Spurious correlationA statistical relationship in training data that doesn't reflect a causal connection — the model learns it and fails when the coincidence disappears.
Adversarial exampleAn input modified (often imperceptibly) to deliberately cause a trained model to make a wrong prediction.
InterpretabilityMethods for understanding which features or patterns a model uses to make its decisions — crucial for catching spurious correlations before deployment.

Lesson 4 Quiz

When Patterns Break — AI Failure Modes
1. The skin cancer AI classifier studied in 2016 achieved high accuracy by learning to detect what spurious feature?
Correct. Dermatologists routinely place rulers next to suspicious lesions — so malignant images contained rulers more often. The AI learned ruler = cancer, which would fail completely in a different photographic context.
The classifier had learned that ruler markings correlated with malignancy — because dermatologists photograph suspicious lesions with rulers for size reference. A brilliant-seeming model had learned a completely wrong pattern.
2. Why did Amazon abandon its AI recruiting tool in 2018?
Exactly right. Ten years of male-dominated hiring decisions were the training data. The model learned that male-associated terms correlated with success — and penalized women's credentials accordingly.
The tool penalized resumes mentioning "women's" organizations and downgraded graduates of all-women colleges — because it trained on ten years of hiring history that skewed heavily male.
3. What did Goodfellow et al. demonstrate with adversarial examples in 2014?
Correct. A panda became a "gibbon" at 99.3% confidence with changes invisible to human eyes — revealing that neural networks learn perceptually alien statistical features, not human-like visual concepts.
Goodfellow's team showed that imperceptible pixel-level noise could make a confident classifier flip entirely — a panda classified as a gibbon with 99% confidence. Neural networks learn patterns humans don't see.
4. What is the core limitation of only measuring a model's accuracy on a held-out test set?
Exactly. The skin cancer classifier and COVID diagnostic AIs all had strong test accuracy while learning the wrong things. High accuracy is necessary but never sufficient — understanding what the model learned is equally critical.
High test accuracy is necessary but not sufficient. Models can learn spurious patterns — rulers, scanner artifacts, resume wording — that produce accurate predictions on test data but fail catastrophically in real deployment.

Lab 4 · AI Failure Detective

Investigate spurious correlations, adversarial examples, and what interpretability tools can reveal.

Your Mission

You've seen how AI can learn with high accuracy while being completely wrong about why. Now probe those ideas. Ask the AI to explain how LIME works, what SHAP values reveal, how physical adversarial patches fool self-driving cars, or how the Amazon recruiting failure could have been caught earlier.

Try asking: "What audit could have caught Amazon's recruiting AI bias before launch?" — or — "How does SHAP help explain a model's decision?"
AI Failure Detective Lab
L4 · Pattern Party

Module 3 · Pattern Party

15 questions · Pass at 80% (12/15)
1. What concept describes the failure mode when real-world data after deployment differs significantly from training data?
Correct. Distribution shift is when the statistical properties of deployment data differ from training data — like Google Flu Trends when the platform's algorithm changed.
This is distribution shift — when real-world data at deployment time differs from what the model trained on, causing its learned patterns to fail.
2. Who created the ImageNet dataset that triggered the deep learning revolution?
Correct. Fei-Fei Li led the creation of ImageNet using Amazon Mechanical Turk workers in 167 countries to label 14 million images — the dataset that enabled AlexNet's 2012 breakthrough.
ImageNet was created by Fei-Fei Li and her Princeton team, using Mechanical Turk workers worldwide to label 14 million images over two years.
3. WebText (used in GPT-3 training) filtered web pages using which mechanism?
Right. Reddit's voting system acted as a quality proxy — but it also baked Reddit's demographic skews into the training data.
WebText used Reddit upvotes as its quality filter — only pages linked from posts with ≥3 upvotes were included.
4. Backpropagation was introduced in its modern form by which team, and in what year?
Correct. The 1986 Rumelhart-Hinton-Williams paper in Nature established backpropagation as the foundational training algorithm for neural networks.
Rumelhart, Hinton, and Williams published their backpropagation paper in 1986 — it became the backbone of modern neural network training.
5. MNIST, created at Bell Labs in 1998, contains how many labeled handwritten digit images?
Correct. MNIST's 70,000 images (60,000 training + 10,000 test) became the standard "hello world" benchmark for supervised classification algorithms.
MNIST contains 70,000 labeled handwritten digit images — 60,000 for training and 10,000 for testing. It became machine learning's standard benchmark.
6. In reinforcement learning, what is a "policy"?
Exactly. A policy maps states to actions — it's what the agent has learned about which actions to take in which situations to maximize long-term reward.
A policy is the learned strategy for choosing actions — mapping environment states to actions in a way that maximizes cumulative reward over time.
7. AlphaGo defeated which European Go champion in 2015, marking the first professional-level computer Go victory?
Correct. AlphaGo beat Fan Hui 5-0 in October 2015 — the first time a computer defeated a professional Go player on a full 19×19 board without handicap.
AlphaGo defeated Fan Hui, the European Go champion, five games to zero in October 2015 — the first such victory over a professional player.
8. OpenAI Five played how many simulated years of Dota 2 in approximately one month?
Correct — 45,000 simulated years of gameplay in a single month, using 256 GPUs and 128,000 CPU cores. Scale of experience is what made its strategy possible.
OpenAI Five played 45,000 simulated years of Dota 2 in about a month — massively parallel self-play at a scale impossible to achieve any other way.
9. RLHF — Reinforcement Learning from Human Feedback — was first described in a paper published by which organization, and in what year?
Correct. OpenAI's 2017 paper introduced RLHF as a way to align agent behavior with human preferences — it later became central to ChatGPT's training.
OpenAI published the foundational RLHF paper in 2017 — the technique became the key ingredient in making large language models like ChatGPT helpful and aligned.
10. The 2016 skin cancer AI study revealed the model was actually detecting what instead of cancer features?
Right. Ruler markings appeared more often in malignant-lesion photos because dermatologists document suspicious lesions with rulers for size reference. The AI learned the wrong pattern entirely.
The model learned that ruler markings correlated with malignancy — a perfect example of a spurious correlation catching a model that looked accurate but had learned the wrong feature.
11. Goodfellow et al.'s 2014 adversarial example paper showed a panda image was reclassified as what, with 99.3% confidence, after imperceptible noise was added?
Correct — gibbon, at 99.3% confidence. The changes were invisible to humans, revealing that neural networks learn perceptually alien statistical features.
The panda was classified as a gibbon with 99.3% confidence after imperceptible pixel noise — demonstrating that neural networks learn statistical features that don't match human visual perception.
12. Amazon's recruiting AI penalized candidates from which educational institutions, among other biases?
Correct. The model penalized "women's" — as in women's clubs — and downgraded graduates of two all-women colleges, having learned that male-associated patterns correlated with successful hiring.
Amazon's tool downgraded graduates of two all-women colleges and penalized resumes mentioning women's organizations — reflecting gender bias baked into a decade of male-dominated hiring decisions.
13. An autoencoder's "bottleneck" forces the model to learn what?
Exactly. The bottleneck constrains the model to compress data into a minimal representation — forcing it to retain only what matters most for reconstruction. Anomalies reconstruct poorly.
The bottleneck forces compression — the model must retain only essential information. This is why autoencoders detect anomalies: unusual inputs that don't fit the learned structure reconstruct poorly.
14. AlphaGo Zero differed from AlphaGo in one fundamental way. What was it?
Correct. AlphaGo Zero started with nothing but the rules of Go and self-play — no human expert moves at all — and surpassed the original AlphaGo within 40 days.
AlphaGo Zero used no human data whatsoever — only Go's rules and millions of self-play games. It surpassed the original AlphaGo, which needed 30 million human expert moves, in just 40 days.
15. The EU's AI Act (proposed 2022) requires conformity assessments for "high-risk" AI systems. Which of these is specifically listed as a high-risk domain?
Correct. High-risk domains under the EU AI Act include medical, employment, and law enforcement — areas where spurious correlations and biased training data can cause serious, irreversible harm.
The EU AI Act designates medical, hiring, and law enforcement AI as high-risk, requiring mandatory bias audits and training data documentation before deployment.