Chapter 6 · Watch, then practise
Machine Learning
Learn model parameters from examples and judge success on unseen data rather than on training performance alone.
3 questions · 3 with related videos. Matches are based on playlist titles; broader background matches are labeled.
What to study
- Features and labels
- Supervised and unsupervised learning
- Training, validation and test sets
- Nonlinear neural networks
Chapter playlists
Choose a playlist
Notes
Supervised Learning In Artificial Intelligence Explained in Hindi
5 Minutes Engineering · 3:32
Choose a video · 2 lectures
The two chapter lectures cover supervised and unsupervised learning, the settings being compared.
1. Learning settings
How do supervised and unsupervised learning differ?
Supervised learning uses examples paired with target outputs, such as messages labeled spam or not spam. Unsupervised learning finds structure without those target labels, such as grouping similar messages. The available feedback determines the learning problem; neither label guarantees that a learned model is useful.
Machine Learning Fundamentals: Cross Validation
StatQuest with Josh Starmer · 6:05
The supplementary cross-validation lesson supports separating model fitting from performance assessment.
2. Generalization
Why separate training, validation and test data?
Fit parameters on training data, use validation data for model choices, and reserve test data for final evaluation. Repeatedly choosing a model based on test scores leaks information from that test. High training accuracy with poor unseen-data accuracy indicates that fitting examples has not generalized well.
Neural Networks Pt. 3: ReLU In Action!!!
StatQuest with Josh Starmer · 8:58
Choose a video · 2 lectures
ReLU and neural-network fundamentals provide the nonlinear-activation concepts used in this answer.
3. Nonlinear models
Why use nonlinear activation functions in a multilayer network?
Composing only linear transformations still produces a linear transformation. Nonlinear activations let hidden layers represent richer relationships. Training adjusts weights to reduce a loss; added capacity also makes evaluation on held-out examples essential. A larger network is not automatically a better predictor.