What is the standard reconstruction loss for autoencoders?
easyAnswer
- MSE (or L1) between input and reconstruction for continuous data.
- Binary cross-entropy per pixel for images normalized to [0, 1] (treats each pixel as a Bernoulli).
- VAEs add a KL divergence between the posterior over latents and a prior — the ELBO objective.
- Modern generative models use perceptual losses (feature-space MSE) or adversarial losses for sharper reconstructions.
Check yourself — multiple choice
- Only cross-entropy
- MSE / L1 (continuous) or pixel-wise BCE ([0,1]) plus KL for VAEs; perceptual/adversarial for sharpness
- Same as attention loss
- Only labels
Reconstruction loss: MSE/L1/BCE (+ KL for VAE); perceptual/adversarial for sharpness.
#losses#generative
Practise Deep Learning
214 interview questions in this topic.
Related questions
- Write the vanilla GAN's minimax objective.
- What is perceptual loss and why is it better than pixel-wise for image quality?
- What does WGAN + gradient penalty change vs vanilla GAN?
- Write the VAE ELBO and explain each term.
- Why compute softmax + cross-entropy jointly via log-sum-exp?
- How should the output head be designed for a regression task with a strictly positive target?