GANs & Autoencoders — Interview Q&A

Answer: GANs consist of two networks: a generator (G) that creates fake data from noise, and a discriminator (D) that tries to distinguish real from fake. They play a minimax game: G tries to fool D, D tries to not be fooled. The equilibrium is when G produces realistic data (D=0.5).

Answer: Generator maps latent vector z (random noise) to data space, trying to produce realistic samples. Discriminator is a binary classifier that outputs probability of input being real. They are trained alternately: D on real + fake, G to maximize D's error.

Answer: Minimax: G minimizes log(1-D(G(z))) → early vanishing gradient. Non-saturating: G maximizes log(D(G(z))) → stronger gradients early. Modern GANs use non-saturating loss with improved training.

Answer: Mode collapse occurs when generator produces limited varieties (only few modes of data distribution). It happens when G finds a few "tricks" that fool D and over-optimizes them, failing to explore full distribution.

Answer: DCGAN (Deep Convolutional GAN) introduced: 1) Replace pooling with strided convolutions (D) / fractional-strided (G). 2) BatchNorm in both G and D. 3) No fully connected layers. 4) ReLU in G (except output tanh), LeakyReLU in D. Stabilized training.

Answer: WGAN replaces JSD with Earth-Mover (Wasserstein) distance, which is continuous and provides meaningful gradients even when D is perfect. Uses weight clipping (later gradient penalty WGAN-GP) for Lipschitz constraint. Solves mode collapse and training instability.

Answer: cGAN feeds additional condition (class label, text, image) to both generator and discriminator. Enables controlled generation. Applications: Pix2Pix, text-to-image synthesis, semantic segmentation.

Answer: CycleGAN uses two generators (G: X→Y, F: Y→X) and two discriminators. Key: cycle-consistency loss – translating X→Y→X should return original. No paired data needed. Also identity loss to preserve color.

Answer: Latent space (z) is low-dimensional input to generator, typically Gaussian. G learns to map continuous z to realistic images; interpolating between z vectors yields semantically smooth transitions, showing G has learned meaningful representations.

Answer: StyleGAN removes input latent vector; instead uses mapping network to intermediate latent space w, then AdaIN (adaptive instance normalization) controls style at each layer. Also adds noise for stochastic variations. Enables disentangled control (coarse/fine styles).

Answer: Inception Score (IS): uses pretrained InceptionNet; measures image quality and diversity (high score if confident class predictions & varied labels). Frechet Inception Distance (FID): computes Wasserstein-2 distance between real & fake feature distributions; lower is better, more robust than IS.

Answer: When D becomes too strong (perfectly classifies), log(1-D(G(z))) saturates to 0, giving G almost no gradient. Solutions: non-saturating loss, WGAN (critic scores not probabilities), label smoothing, or making D weaker.

Answer: Replace real labels (1) with soft values like 0.9. Prevents D from becoming overconfident, providing smoother gradients. Only smooth real labels; smoothing fake labels (0→0.1) encourages D to push G samples away, harming training.

Answer: Weight clipping forces critic to lie in narrow space, leading to capacity underuse and exploding/vanishing gradients. WGAN-GP replaces it with gradient penalty: penalize if gradient norm deviates from 1 (Lipschitz constraint).

gp = lambda * ((grad_norm - 1)**2).mean()

Answer: Normalizes weights by their largest singular value, enforcing Lipschitz constraint (spectral norm = 1). Used in SNGAN; stabilizes training without heavy hyperparameter tuning. Works well for both G and D.

Answer: SAGAN uses self-attention to model long-range dependencies (global features) instead of only local convolutions. Improves image quality in complex scenes (e.g., ImageNet) by capturing relationships between distant regions.

Answer: G is trained to match the expected features (intermediate activations) of real data from D, not just final D output. Minimizes L2 distance between real/fake feature means. Helps prevent overtraining on current D.

Answer: Start training with low-resolution images, gradually add layers to increase resolution. Stabilizes high-resolution GAN training (e.g., 1024x1024). Both G and D grow simultaneously. Used in StyleGAN, ProGAN.

Answer: Nash equilibrium: D is optimal (cannot distinguish real/fake) and G is optimal (data distribution = real distribution). In practice, GANs oscillate and rarely converge to exact equilibrium; we aim for approximate Nash. Techniques like consensus optimization try to find stable points.

Answer: An autoencoder is an unsupervised neural network that learns to copy its input to output via a bottleneck (latent) layer. Architecture: Encoder compresses input to latent code; Decoder reconstructs from latent code. Trained with reconstruction loss (e.g., MSE).

Answer: Undercomplete: bottleneck dimension less than input dimension. Forces compression, learns useful features. Overcomplete: bottleneck dimension larger than input. Risks learning identity function; requires regularization (sparse, denoising).

Answer: MSE for continuous values, binary cross-entropy for pixel values in [0,1] (treating as probabilities), L1 loss for robustness.

Answer: Dimensionality reduction (PCA nonlinear), anomaly detection (high reconstruction error), denoising, inpainting, compression, generative modeling (VAE), feature extraction for pretraining.

Answer: Input is corrupted (e.g., add noise, mask), model learns to reconstruct clean original. Forces learning robust features, not just copying. Corrupting process: x̃ = x + ε, ε ~ N(0, σ²).

Answer: Enforces that most latent units are inactive. Sparsity penalty added to loss: KL divergence between average activation and target (e.g., ρ=0.05). Also L1 regularization on activations. Encourages specialized feature detectors.

Answer: CAE adds penalty: Frobenius norm of Jacobian of encoder ∂z/∂x. Encourages robustness to small input changes by making latent representation contractive. DAE corrupts input; CAE regularizes gradient.

Answer: VAE is a generative model that learns latent distribution p(z|x). Encoder outputs parameters of Gaussian (μ, σ). Decoder generates data from sampled z. Trained with ELBO: reconstruction + KL divergence to prior p(z)=N(0,I). Enables interpolation and generation.

Answer: Sampling from N(μ, σ) is non-differentiable. Trick: sample ε ~ N(0,1), then z = μ + σ·ε. Gradient flows through μ,σ, enabling backprop.

z = mu + sigma * torch.randn_like(sigma)

Answer: AE learns deterministic latent code; VAE learns probabilistic latent distribution. AE latent space not continuous; VAE enforces smoothness via KL. VAE is generative; AE is reconstructive.

Answer: VAE variant with β > 1 multiplier on KL term. Encourages more disentangled latent representations. Trade-off: reconstruction vs. independence.

Answer: Train on normal data only. Anomalies yield high reconstruction error. Set threshold based on validation. Used in fraud, industrial defect detection.

Answer: For image data. Encoder uses Conv+Pooling, decoder uses Transposed Conv/Upsampling. Preserves spatial structure.

Answer: Multiple autoencoders stacked; each layer trained separately to reconstruct previous layer's output. Used for deep network pretraining (historical). End-to-end fine-tuning.

Answer: Sparse penalty (KL, L1), denoising (corrupt input), contractive (Jacobian), variational (KL to prior), dropout.

Answer: Linear autoencoder (no nonlinearity) with MSE learns the same principal subspace as PCA. Weights span the same space, but not necessarily orthogonal.

Answer: Decoder ignores latent z, and q(z|x) matches prior p(z). Happens with strong decoders (e.g., autoregressive). Solutions: KL annealing, free bits, β-VAE, reducing decoder capacity.

Answer: AAE uses adversarial training to match aggregated posterior of latent code to prior. Discriminator distinguishes true prior samples vs encoder output. Enables more flexible priors.

Answer: For images, visualize decoder weights or generate from latent traversal. For latent space, t-SNE/UMAP on z. For VAE, interpolate between z1 and z2.

Answer: Yes: train denoising autoencoder with random masking. Model learns to impute missing values. Also partial autoencoders. VAE can model conditional distribution.