AI Researchers Propose Neural Diffusion Processes (NDPs), A Novel Approach Based Upon Diffusion Models, That Learns To Sample From Distributions Over Functions

Neural Diffusion Processes (NDPs), a proposed denoising diffusion model approach for learning probabilities on function spaces and generating prior and conditional function samples. NDPs generalize diffusion models to function spaces with an infinite number of dimensions by permitting the indexing of random variables. They demonstrate that this model can capture functional distributions similar to the true Bayesian posterior. The bi-dimensional attention block is a new component for building neural networks that wire size and sequence equivariance into the neural network architecture to behave like a stochastic process.

Traditionally, researchers have used Gaussian processes (GP) to specify prior and posterior distributions over functions. However, this approach becomes computationally expensive when scaled, is constrained by the expressivity of its covariance function, and struggles to accommodate a point estimation for the hyperparameters.

In the new paper Neural Diffusion Processes, a research team addresses these issues, proposing Neural Diffusion Processes. The novel framework learns to sample from rich distributions over functions at a lower computational cost and to capture distributions close to the true Bayesian posterior of a standard Gaussian process.

The article explains that Bayesian inference for regression is advantageous but frequently expensive and requires a priori modeling assumptions. This is the presumption underlying neural diffusion processes.

The group summarizes their key contributions as follows:

  • The team proposes a novel model, the Neural Diffusion Process (NDP), which extends the application of diffusion models to stochastic processes and is capable of describing a diverse distribution over functions.
  • The team takes special care to incorporate the known symmetries and properties of stochastic processes, such as exchangeability, into the model, which facilitates the training procedure.
  • The team demonstrates the capabilities and adaptability of NDPs by applying them to various Bayesian inference tasks, such as prior and conditional sampling, regression, hyperparameter marginalization, and Bayesian optimization.
  • The team also presents a novel method for global optimization using NDPs.

The proposed NDP is a denoising diffusion model-based method for learning probabilities from a function and generating prior and conditional function samples. It permits complete marginalization over the GP hyperparameters while reducing the computational load compared to GPs.

The team analyzed the sample quality of existing cutting-edge neural network-based generative models. Based on their findings, they developed NDP to generalize diffusion models to infinite-dimensional function spaces by permitting the indexing of random variables onto which the model diffuses.

The researchers also adopted a novel bi-dimensional attention block to ensure input dimensionality and sequence equivariance and to permit the model to draw samples from a stochastic process. Therefore, NDP can capitalize on stochastic processes, such as exchangeability.


In their empirical study, the team assessed the ability of the proposed NDP to generate high-quality conditional samples, marginalize over kernel hyperparameters, and maintain input dimensionality invariance.

The results demonstrate that NDP can capture functional distributions close to the true Bayesian posterior while simultaneously reducing computational demands.

The researchers notice that while the number of diffusion steps improves NDP sample quality, it also slows inference times. According to the authors, future research could investigate inference acceleration or sample parameterization techniques to address this issue.

Future Work- Researchers discovered, as with other diffusion models, that the sample quality of an NDP increases with the number of diffusion steps T. This results in slower inference times compared to different architectures like GANs. To address this issue, techniques for accelerating the inference process could be implemented. The method proposed by Watson et al. is of particular interest to NDPs because it provides them with a moral and distinct metric to assess sample quality, namely the marginal likelihood of the corresponding GP. In conclusion, parameterizing the samples in the Fourier domain might be an intriguing alternative method.

This Article is written as a summary article by Marktechpost Staff based on the paper 'NEURAL DIFFUSION PROCESSES'. All Credit For This Research Goes To Researchers on This Project. Checkout the paper, code (coming soon), ref article.

Please Don't Forget To Join Our ML Subreddit