Unsupervised learning finds patterns in data without relying on labels that supply the correct answer for each example. This lesson introduces the learning framework, the kinds of data it uses, and how unsupervised methods relate to generative and discriminative tasks.
In unsupervised learning, an algorithm learns patterns from data that has no assigned target labels. Rather than being told that an image contains a cat or that a message is spam, the algorithm looks for structure in the examples themselves.
All flashcards
What is unsupervised learning?
A machine-learning framework in which an algorithm learns patterns or structure from data without supplied target labels.
How does self-supervised learning obtain a training signal?
It constructs targets from the data itself, such as asking a model to predict a portion of an example that was hidden.
What does an energy function represent in a Boltzmann machine?
It summarizes the energy of network activation states; lower-energy states receive higher probability.
What is Hebbian learning?
A learning principle in which connections strengthen based on coincident neural activity.
What does a variational autoencoder's encoder represent?
A distribution over latent variables given an input.
Quiz questions
In a Boltzmann machine, how does the probability of an activation state relate to its energy?Answer: States with lower energy are assigned higher probability.
How can a Hopfield network function as a content-addressable memory?Answer: A partial or noisy cue can prompt recall of a stored pattern.
What limitation can affect the expectation–maximization algorithm when learning a latent-variable model?Answer: It can settle at a local optimum rather than recover the true parameters.
What property does a self-organizing map aim to preserve in its topographic map?Answer: Inputs with similar properties tend to occupy nearby map locations.
What is the main role of a sigmoid belief network's neurons?Answer: They act as binary stochastic units in a directed probabilistic network.