Diffusion models are used in AI to generate high-quality samples from complex data distributions, and discriminative diffusion models are used for tasks like classification or regression. These models offer advantages like better handling of uncertainty, robustness to noise, and capturing complex data dependencies.
Generative models can identify anomalies in data by quantifying the deviation of a new data point from the learned data distribution, aiding in anomaly detection tasks. Instead of being competitive alternatives, researchers at Carnegie Mellon University couple these two models during the inference stage to leverage the benefits of both generative and discriminative models.
The team built a Diffusion-based Test Time Adaptation (TTA) model that adapts methods from image classifiers, segmenters, and depth predictors to individual unlabelled images. Their model is reminiscent of an encoder-decoder architecture: a pre-trained discriminative model encodes the image into a hypothesis used as conditioning for a pre-trained generative model to generate the image.
Diffusion-TTA effectively adapts image classifiers for in- and out-of-distribution examples across established benchmarks and fine-tunes the model using image reconstruction loss. It consistently outperforms Diffusion Classifier and is effective across multiple discriminative and generative diffusion model variants.
Researchers also present an ablative analysis to study the effect of adapting different model parameters. They conclude that by using generative models, users can obtain efficient results by co-training the Diffusion-TTA model under a joint discriminative task loss and a self-supervised image reconstruction loss.
For more information, you can check out the Paper and Project. All credit for this research goes to the researchers of this project. Also, don’t forget to join the ML SubReddit, Facebook Community, Discord Channel, and Email Newsletter, where they share the latest AI research news, cool AI projects, and more.
If you like their work, you’ll love their newsletter.