Score-Based Generative Models Quiz

Answers and explanations appear as you take the quiz.

1. What exactly is the 'score' that these models learn?
   • The gradient of the log-probability density with respect to the input
   • A quality rating from 0 to 100 for each image
   • The number of training epochs completed
   • The classification label of the data point
2. Why are score-based models trained on data corrupted with noise at multiple scales?
   • To make the images look artistic
   • Because the score is hard to estimate where clean data is sparse, and noising fills in those regions
   • To reduce the file size of the dataset
   • Because noise speeds up the GPU
3. Which sampling procedure did early score-based models use to generate data?
   • Gradient descent on the pixels toward a fixed target image
   • A single matrix multiplication
   • Langevin dynamics — stepping along the score while adding noise
   • Random cropping and flipping
4. What major connection did the 2021 score-SDE paper establish?
   • That GANs are always better than diffusion
   • That neural networks cannot model images
   • That score models only work in two dimensions
   • That diffusion models and score-based models are two views of the same continuous SDE process
5. How does predicting noise relate to predicting the score in denoising score matching?
   • The conditional denoising-score target is −ε/σ for standardized Gaussian noise; the model estimates the noisy-data score
   • Noise prediction and score estimation are unrelated
   • Predicting noise is impossible
   • The score is always zero for noised data
6. What does the deterministic 'probability flow ODE' enable that pure stochastic sampling does not easily provide?
   • Higher GPU power consumption
   • Exact likelihood computation and a deterministic sampling path
   • Automatic dataset labeling
   • Removal of the need for any neural network
7. In score-based generation, what does the reverse-time SDE do?
   • It converts image labels into captions
   • It adds noise to an already generated output only
   • It transforms a prior noise distribution back toward the data distribution
   • It estimates the class label of each pixel
8. In a score-based MRI reconstruction method, how can a learned image prior help recover an undersampled scan?
   • Use the prior alone and ignore measured k-space data
   • Combine the learned score prior with the measurement constraint during posterior sampling
   • Replace all measurements with random noise
   • Treat the score as a diagnosis of the patient