Short answers, pulled from the story.
Unsupervised learning is a framework in machine learning where algorithms learn patterns exclusively from unlabeled data, without any human-supplied classifications or tags. It contrasts with supervised learning, where datasets are manually constructed and labeled. Common applications include clustering, anomaly detection, and learning latent variable models.
Supervised learning trains on manually labeled datasets, such as ImageNet1000, which are expensive to construct. Unsupervised learning uses unlabeled data harvested cheaply at scale, such as large web text corpora gathered via web crawling with only minor filtering. Supervised learning infers a conditional probability distribution given a label; unsupervised learning infers a prior probability distribution from the data itself.
Key unsupervised neural networks include the Hopfield network, Boltzmann machine, Restricted Boltzmann Machine, Helmholtz machine, and Variational Autoencoder. The Hopfield network was described by John Hopfield in 1982. The Boltzmann machine was introduced by Hinton and Sejnowski in 1983. Variational Autoencoders were introduced by Kingma, Rezende, and colleagues in 2013.
Geoffrey Hinton and Terry Sejnowski introduced the Boltzmann machine in 1983, following earlier work by Sherington and Kirkpatrick from 1975. Its energy function is inspired by Ludwig Boltzmann's analysis of a gas's macroscopic energy derived from the microscopic probabilities of particle motion.
Hebbian learning is the classical example of unsupervised learning in neural networks, based on Donald Hebb's principle that neurons that fire together wire together. Connection strength is reinforced by coincident neuron activity, not by error feedback. It has been proposed as an explanation for cognitive functions including pattern recognition and experiential learning.
The method of moments is a statistical approach where unknown model parameters are estimated by relating them to the moments of random variables computed from data. First-order moments are mean vectors and second-order moments are covariance matrices; higher-order moments are represented as tensors. Unlike the Expectation-Maximization algorithm, the method of moments guarantees global convergence under certain conditions.