Generative models, a class of probabilistic machine learning, have seen extensive use in various fields, such as the visual and performing arts, medicine, and physics. These models are proficient in creating probability distributions that accurately describe datasets, making them ideal for generating synthetic datasets for training data and discovering latent structures and patterns in an unsupervised learning environment.
A particular category of generative models, known as diffusion models, is built through a two-step process: the forward and inverse processes. During the forwarding process, the data distribution corrupts over time, veering from its pristine state to a noisy one. Concurrently, the reverse process can restore the data distribution by learning to reverse the corruptions introduced earlier. While diffusion models have shown excellent performance, they traditionally employ a fixed forward process that hinders task adaptation or target simplification during the reverse process.
A groundbreaking study conducted by researchers at the University of Amsterdam and Constructor University, Bremen, has introduced a novel framework named Neural Flow Diffusion Models (NFDM). Unlike traditional diffusion models, NFDM allows the forward process to learn latent variable distributions, provided any continuous distribution can be represented as an invertible mapping applied to noise. Through an end-to-end optimization technique, the researchers have minimized a variational upper bound on the negative log-likelihood (NLL). Further, they have proposed a parameterization for the forward process based on efficient neural networks, facilitating easier learning of data distribution and better adaptability during its training.
By leveraging NFDM’s adaptability, the team explored training with restrictions on the reverse process to understand generative dynamics with specific attributes. The experiments yielded improved computing efficiency compared to baselines on synthetic datasets, MNIST, CIFAR-10, and downsampled ImageNet.
When testing NFDM on CIFAR-10 and ImageNet 32 and 64, the researchers demonstrated the model’s extensive potential with a learnable forward process. The impressive NLL results achieved have essential implications for various applications like data compression, anomaly detection, and out-of-distribution detection. The NFDM also exhibited superior capabilities in learning generative procedures with specific attributes such as dynamics with straight-line trajectories, enabling faster sampling rates, bettering generation quality, and reducing the required sampling steps.
Nonetheless, implementing NHDM has its challenges, and the researchers caution that its use typically increases computational costs compared with conventional diffusion models. The data indicates that NFDM optimization iterations tend to take around 2.2 times longer than traditional models. However, the researchers believe that NFDM’s flexibility in learning generative processes and its potential across various applications justifies these challenges. They also suggest prospects for further enhancement, like incorporation of orthogonal methods, target alteration, and exploring different parameterizations.
