Denoising Diffusion Models

So far, AI has been all about giving solutions. Is this an African or a European swallow? Will this drug work on this disease or not? But knowing only the direct answer to these questions, does not teach us anything new about the world. On the other hand, if a model could tell us why this is a European swal-low and not an African swallow or why this drug will work on this disease, then we have not only been told the solution, but how to arrive at the solution, and this might lead to some truly new insight. We might be able to tell the swallows apart the next time we encounter them, or we might think of other drugs which will also work on the disease.

Denoising diffusion models are a kind of generative model which are trained by diffusing data and then learning how to denoise it. This means that these models can create new data from complete noise. They have been used successfully for many generation tasks, like making images, synthesizing speech and simulating protein folding. However, they still have several problems. They take a long time to train, they take a long time to sample from, and if you generate an image from them using a prompt asking for a specific bird, you might get your solution, but you have no way of asking why this is an image of the bird you asked for. Since a denoising diffusion model can be used to generate both images of African and European swallows, it must contain information on what these swallows look like and how they are different, since we can ask for one and not the other, but how do we make the model tell us in a way that we can understand?

My project is about improving diffusion models in two ways. Making them generate better samples, or the same quality but faster, and find a way to let us question the model about why the generated sample satisfies the prompt.
Hopefully, this will be one small step to a future where AI models will be able to explain to humans what they are doing.