trevorkang / ebm_jem

Assignment 3, Advanced Deep Learning @ FAU

EBM and JEM

EBM (Energy-Based Model)

Basic Assumption: Higher the energy, lower the probability.

JEM (Joint Energy-Based Model)

Generated samples are used to estimate the partition function.

Final Example:

Training Strategies for Stability

Contrastive Divergence Loss

Reservoir Sampling - SGLD :

MCMCSampler.synthesize_samples(): perform SGLD sampling for a given number of steps, and return the final samples. Set up different replay buffers for conditional and unconditional sampling.

The sampling process is shown below: class label=7

Out-of-Distribution (OOD) Detection

run_ood_detection(): OOD detection on GLYPHS dataset. We use two extra datasets to test the performance.

score_fn(): score function for OOD detection. Use the negative log-likelihood p(x) as the score function.

The histogram of the score function is shown below:

JEM can be used as a binary classifier:

0: in-distribution (ID): all blue bars

1: out-of-distribution (OOD): all orange bars

after setting the threshold (0.25), we can use it to detect OOD.

Corresponding ROC score : 0.9153815771938624