The application of diffusion models in data science

Abstract

The diffusion model is a versatile mathematical model that has been applied in various felds. The diffusion model represents decision making as a stochastic process where the decision maker accumulates evidence over time by integrating noisy inputs. The model assumes that the input signals are normally distributed, and that the decision maker has to decide which of two possible choices the signal belongs to. The decision maker must continuously accumulate evidence until a decision boundary is reached. Once the evidence reaches a threshold, a decision is made. We have studied the logic underlying diffusion models, including the denoising diffusion probabilistic model (DDPM), the denoising diffusion implicit model (DDIM), and the latent diffusion model. The latent diffusion model can achieve comparable performance to the diffusion model while using fewer parameters and less computation. It is also easier to optimize, and can often produce higher quality results for image denoising and generation tasks. Additionally, We studied three applications of the diffusion model: Unconditional image generation, Text-to-Image and Inpainting. For Unconditional image generation, we trained the model on a diabetic retinopathy dataset. The images generated by these applications prove that diffusion models can produce high quality pictures and become useful worldwide.