Autoencoders, VAEs, GAN training dynamics, diffusion models, latent spaces and evaluation metrics like FID.
34 cards · basic cards · AI-written, checked twice. Edit anything.
- What is the main goal of an autoencoder?
Show answer
To learn a compressed representation of input data by training an encoder to map inputs to a latent code and a decoder to reconstruct the input from that code.- What are the two main components of an autoencoder architecture?
Show answer
The encoder, which compresses the input into a latent representation, and the decoder, which reconstructs the input from the latent code.- What is the bottleneck in an autoencoder?
Show answer
The narrow layer at the middle of the network where the encoder outputs a compressed representation, forcing the network to learn compact encodings.- What loss function is commonly used to train autoencoders?
Show answer
Reconstruction loss, typically mean squared error (MSE) or binary cross-entropy between the input and reconstructed output.- Name two applications of autoencoders.
Show answer
Dimensionality reduction, anomaly detection, denoising, feature learning, or data compression.- How does a VAE differ from a standard autoencoder?
Show answer
A VAE adds a probabilistic encoder that outputs a distribution (mean and variance) over the latent space, plus a KL divergence term in the loss to regularize the latent distribution.- What is the reparameterization trick in VAEs?
Show answer
A technique to sample from a learned distribution by computing z = mu + sigma * epsilon, where epsilon is standard normal, allowing gradients to flow through stochastic sampling.- What is the ELBO in VAE training?
Show answer
The Evidence Lower Bound, the training objective balancing reconstruction and regularization: ELBO = E[log p(x|z)] - KL(q(z|x) || p(z)).- What does the KL divergence term in the VAE loss accomplish?
Show answer
It regularizes the posterior distribution q(z|x) to stay close to the prior distribution p(z), typically a standard normal, making the latent space smooth.- What distribution does a VAE encoder typically output?
Show answer
A diagonal Gaussian distribution, parameterized by a mean vector and diagonal covariance (or log-variance) vector.- How does the VAE prior enable generation?
Show answer
The prior p(z) is typically a standard normal, so samples from this prior can be decoded to generate new data without using an encoder.- What is posterior collapse in VAEs?
Show answer
A failure mode where the learned posterior q(z|x) becomes identical to the prior, causing the encoder to be ignored and the latent code to be unused.- Name one advantage of VAEs over GANs.
Show answer
VAEs provide a principled probabilistic framework with stable training and the ability to compute or bound data likelihood.- What are the two networks in a GAN?
Show answer
The generator, which creates fake samples from noise, and the discriminator, which classifies inputs as real or fake.- What is the adversarial loss objective in GANs?
Show answer
The generator minimizes the probability that the discriminator correctly classifies its samples as fake, while the discriminator maximizes correct classification of real and fake samples.
FlashKeepers