11/11 This is joint work with @willberghammer, @haoyu_wang66, @EnnemoserMartin, @HochreiterSepp, and @sebaleh. See you at #ICLR!
[Poster Link](iclr.cc/virtual/202...)
[Paper Link](arxiv.org/abs/2502.08696)
---

10/11 🏆 Our method outperforms autoregressive approaches on Ising model benchmarks and opens new avenues for applying diffusion models to a wide range of scientific applications in discrete domains.

9/11 📊 Due to the mass-covering property of the fKL, it excels at unbiased sampling. Conversely, the rKL is mode-seeking, making it ideal for combinatorial optimization (CO) as it achieves better solution quality with fewer samples.

8/11 💡 𝐒𝐨𝐥𝐮𝐭𝐢𝐨𝐧 2: We address the limitations of the fKL by combining it with Neural Importance Sampling over samples from the diffusion sampler. This allows us to estimate the gradient of the fKL using Monte Carlo integration, making training more memory-efficient.

7/11 An alternative is the forward KL divergence (fKL), where it is well known how to increase memory efficiency by leveraging Monte Carlo integration over diffusion time steps. However, the fKL divergence requires samples from the target distribution!

6/11 💡 𝐒𝐨𝐥𝐮𝐭𝐢𝐨𝐧 1: We apply the policy gradient theorem to the rKL between joint distributions of the diffusion path. This enables the use of mini-batches over diffusion time steps by leveraging reinforcement learning methods, allowing for memory-efficient training.

5/11 A commonly used divergence is the reverse KL divergence (rKL), as the expectation of the divergence goes over samples from the generative model. However, naive optimization of this KL divergence requires backpropagating through the whole generative process.

4/11 🚨 𝐂𝐡𝐚𝐥𝐥𝐞𝐧𝐠𝐞: However, existing diffusion samplers struggle with memory scaling, limiting the number of attainable diffusion steps due to backpropagation through the entire generative process.

3/11 🔍 𝐃𝐢𝐟𝐟𝐮𝐬𝐢𝐨𝐧 𝐒𝐚𝐦𝐩𝐥𝐞𝐫𝐬 aim to sample from an unnormalized target distribution without access to samples from this distribution. They can be trained by minimizing a divergence between the joint distribution of the forward and reverse diffusion paths.

2/11 We've developed scalable and memory-efficient training methods for diffusion samplers, achieving state-of-the-art results in combinatorial optimization and unbiased sampling on the Ising model.

A Pizza Steel or Pizza Stone with max Heat (250 celsius) should do the Job

I think it is fine to keep the score, but if all concerns are addressed they should at least justify why they are nevertheless keeping their score.

Does this mean all Paper at 6 or above should be accepted?

That is a cool idea!

✋

I also would like to join :)