Lecture 13 - Unsupervised Learning Flashcards
What is unsupervised learning?
Unsupervised learning involves discovering patterns within data without labeled examples, focusing on describing variability in the data and estimating distributions.
What is generative modeling in the context of unsupervised learning?
Generative modeling aims to approximate high-dimensional distributions of data, allowing for the generation of new data samples that resemble the original dataset.
Explain unsupervised representation learning.
Unsupervised representation learning involves learning useful weights and representations from unlabeled data, which can be transferred to other tasks with minimal labeled examples.
What is the main idea behind self-supervised learning?
Self-supervised learning creates surrogate tasks that do not require labels, enabling the model to learn representations by generating labels from the data itself and optimizing a pretext task.
Describe the concept of an autoencoder.
An autoencoder is a neural network that learns to encode input data into a lower-dimensional representation and then decodes it back to the original data, minimizing reconstruction error.
What is a variational autoencoder (VAE)?
A VAE is a type of autoencoder that uses variational inference to encode input data into a probabilistic latent space and decode it, allowing for both data reconstruction and generation.
Explain the role of the Evidence Lower Bound (ELBO) in VAEs.
ELBO is a variational bound used to approximate the likelihood of the data, allowing the optimization of the VAE by maximizing this bound instead of the intractable true likelihood.
What is a Generative Adversarial Network (GAN)?
A GAN consists of two neural networks, a generator and a discriminator, that compete in a minimax game, with the generator creating realistic data samples to fool the discriminator, which tries to distinguish between real and fake samples.
Describe the concept of contrastive learning.
Contrastive learning involves training models to distinguish between similar and dissimilar pairs of data, encouraging the model to learn representations that bring similar samples closer and push dissimilar samples apart.
What is the purpose of a latent variable model?
Latent variable models assume that data lies in a lower-dimensional subspace and build a mapping between the high-dimensional data space and this latent space, facilitating tasks like data generation and reconstruction.
Explain the main idea behind flows in unsupervised learning.
Flows use invertible transformations to map between data space and latent space, allowing for exact likelihood computation and sampling without the need for approximation.
What is the challenge with high-dimensional data in unsupervised learning?
High-dimensional data requires more samples to effectively learn and model the underlying distribution, making it computationally challenging and necessitating advanced techniques to handle the dimensionality.
Write the formula for the reconstruction loss in an autoencoder.
Provide the formula for the variational lower bound (ELBO) in VAEs.
What is the formula for the contrastive loss?