🤖 Machine Learning Fundamentals

What you’ll learn

1. The Shift: From Rules to ExamplesEstablish the founding idea of machine learning: learning patterns from labeled examples instead of writing explicit rules.Traditional programming requires a human to write the rules, which fails for fuzzy perception tasks like vision and speech. Machine learning inverts this: you supply labeled examples and an algorithm derives the rules itself. The same shift already powers spam filters, speech-to-text, face unlock, and much of what a robotaxi perceives on the road.
2. What a Model Actually IsDefine features, labels, and a model as a function with tunable weights.An example is a set of features (measurable inputs) paired with a label (the answer to predict). A model is a mathematical function whose internal numbers (weights) can be adjusted like knobs. Learning means automatically turning those knobs until the model's outputs match the known answers.
3. Training: How the Knobs Get TurnedExplain loss and gradient descent as the mechanism of training.Training reframes learning as minimizing a single error number, the loss. Gradient descent finds the weight settings with the lowest loss by repeatedly computing the slope of the error and taking a small step downhill. The learning rate sets the step size — too big overshoots, too small is slow.
4. The Three Ways Machines LearnDistinguish supervised, unsupervised, and reinforcement learning by the feedback they use.Supervised learning uses labeled examples to predict labels for new inputs. Unsupervised learning has no labels and finds hidden structure like clusters. Reinforcement learning has no dataset but an environment that gives rewards. Deep Blue's 1997 chess win was brute-force search, not this kind of learning.
5. Overfitting: Memorizing Is Not LearningExplain overfitting and the train/test split as the discipline that detects it.A model with too much freedom can memorize its training examples instead of learning the general pattern, scoring high on seen data and failing on new data. Data is split into a training set and an unseen test set; only the test-set score predicts real-world performance.
6. Neural Networks and Deep LearningDescribe neural networks as layered simple units and explain what 'deep' means.A neural network wires many simple units into layers; each layer builds more abstract features, from edges to parts to whole objects. 'Deep' means many stacked layers. Deep networks have millions of weights and took off in the 2010s once large labeled datasets and GPUs supplied the needed data and computing power.
7. Garbage In, Bias OutExplain how bias in training data produces biased models, and why data quality matters more than algorithm choice.Because a model learns the pattern in its examples, biases in the data are learned and applied at scale. Documented cases include Amazon's recruiting tool (Reuters, 2018) and Gender Shades (Buolamwini and Gebru, 2018): up to 34.7% error on darker-skinned women versus 0.8% on lighter-skinned men. The fix is usually better, more representative data and testing across groups.
8. A Short History and Where It Is GoingTrace key milestones and show that data and computing power, not just algorithms, drove progress.Key ideas are old: Turing's 1950 question, Rosenblatt's 1957–60 Perceptron, backpropagation popularized in 1986, Deep Blue in 1997 (search, not learning), AlexNet's 2012 ImageNet result (15.3% vs 26.2%), and the 2017 Transformer. Progress repeatedly waited for enough data and computing power to catch up to existing ideas.

Questions this course answers

Grounded in trusted sources

• Tom M. Mitchell, Machine Learning (McGraw-Hill, 1997) — the standard definition of learning from experience on a task, measured by a performance score
• Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton, 'ImageNet Classification with Deep Convolutional Neural Networks,' NeurIPS 2012 — ILSVRC-2012 top-5 error 15.3% vs 26.2% runner-up: https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks
• David E. Rumelhart, Geoffrey E. Hinton, and Ronald J. Williams, 'Learning representations by back-propagating errors,' Nature 323 (1986): 533–536: https://doi.org/10.1038/323533a0
• Ashish Vaswani et al., 'Attention Is All You Need,' NeurIPS 2017: https://arxiv.org/abs/1706.03762
• Joy Buolamwini and Timnit Gebru, 'Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification,' PMLR 81 (2018): https://proceedings.mlr.press/v81/buolamwini18a.html
• A. M. Turing, 'Computing Machinery and Intelligence,' Mind 59, no. 236 (1950): 433–460
• Frank Rosenblatt, 'The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain,' Psychological Review 65, no. 6 (1958): 386–408; and Cornell Aeronautical Laboratory Report 85-460-1 (1957)
• IBM, 'Deep Blue' — ~200 million chess positions per second, brute-force search, May 1997 match: https://www.ibm.com/history/deep-blue

Every Wunder lesson is built from real, reputable sources — never invented.

Related Science courses