TL;DR
The Sequential Score Sampler (SSS) introduces a "corrected self-training" framework that allows neural diffusion models to learn complex, multimodal distributions without needing pre-existing target samples. By combining a temperature ladder, path-space importance sampling, and replay buffers, SSS achieves state-of-the-art sampling accuracy with significantly lower computational overhead than previous amortized samplers.
What Problem Does This Paper Address?
Sampling from un-normalized distributions—where we know the energy function $E(x)$ but not the normalizing constant—is a cornerstone of physics, chemistry, and statistics. However, two major hurdles remain:
- Multimodality: Traditional MCMC methods often get stuck in local energy minima, failing to jump between isolated "basins" of high probability.
- Computational Cost: While methods like Sequential Monte Carlo (SMC) can handle multimodality, they are expensive to run every time you need new samples.
"Amortized" samplers aim to solve the second problem by training a generative model (like a normalizing flow or diffusion model) to act as a fast, reusable proposal. The catch? Training these models usually requires representative samples from the target distribution—the very thing we are trying to find in the first place.
Methods
The authors propose a corrected self-training scheme that breaks the "chicken and egg" cycle of training samplers. The core workflow is summarized in Figure 1:
- Temperature Ladder: Instead of jumping straight to the complex target, the model learns a sequence of distributions ranging from a simple Gaussian ($\beta=0$) to the target ($\beta=1$).
- Corrected Proposals: The model generates samples at each temperature. These are corrected using importance weights to account for the mismatch between the model's current belief and the true target.
- MCMC Rejuvenation: Resampled particles undergo local MCMC steps to diversify the population and explore nearby configurations.
- Replay Buffers: Weighted samples are stored in buffers ($D_0, \dots, D_K$). This allows the model to reuse costly sampling results over multiple optimization steps, stabilizing the learning process.
- Sequential Score Sampler (SSS): This is the specific implementation using score-based diffusion. A single neural network is trained to predict the "relative score" (the gradient of the log-density relative to a Gaussian) across all temperatures simultaneously.
Datasets
The authors evaluated SSS on five challenging benchmarks:
- GMM16 & GMM50: Gaussian Mixture Models with $2^{16}$ and $2^{50}$ modes, respectively, testing the ability to recover massive numbers of isolated modes.
- DW-4: A 4-particle double-well system, a standard benchmark for interacting particles.
- LJ-13: A 13-particle Lennard-Jones cluster involving steep repulsive forces and complex metastable states.
- Alanine Dipeptide: A real-world molecular system (22 atoms) where the goal is to capture backbone angle conformations at 300K.
Key Insights and Ideas
- Mode Balancing via Importance Sampling: A critical finding is that MCMC alone isn't enough to fix the "relative masses" of different modes. Importance sampling on the diffusion path space is what ensures the model learns the correct global structure.
- Joint Training Across Temperatures: By training one model on the entire temperature ladder, features learned at easier (higher temperature) levels help the model generalize to the harder target level.
- The Power of the Buffer: The paper highlights that the "final-buffer" samples—those that have already been corrected and rejuvenated during training—often provide the most accurate representation of the target, even compared to fresh importance-corrected samples from the trained model.
Key Results
SSS consistently outperformed or matched recent baselines like PITA, TA-BG, and Adjoint Sampling (AS) across various metrics (Distance-$W_2$, Energy-$W_2$, and Geometric-$W_2$).
- Accuracy: On GMM50, importance correction reduced the energy error from 4.304 to 0.321.
- Efficiency: SSS is remarkably fast. On the DW-4 benchmark, SSS took only 15 minutes to train and sample, compared to 530 minutes for PITA.
- Molecular Modeling: For Alanine Dipeptide, SSS buffer samples achieved the lowest energy discrepancies (Energy-$W_2$ of 0.257 vs. 1.615 for PITA).
Limitations and Future Work
- Initial Coverage: The method depends on the training buffers having "adequate coverage." If the initial SMC pass misses a mode entirely, the self-training loop may never discover it, potentially reinforcing the gap.
- Amortization Trade-off: The initial cost of setting up the temperature ladder and training the model is significant. The authors note that the value of amortization depends on how many samples you ultimately need; for a single one-off sampling task, a long-run SMC might still be more efficient.
- Future Directions: Exploring how to better discover missing modes during the self-training process and applying the method to even larger molecular systems are natural next steps.
Summary updated