Is diffusion model supervised learning or unsupervised?

[Alan_He]

At the end of wk1 video, Andrew mention that diffusion model is supervised learning by heart. However, diffusion model is unsupervised learning if you google it or ask chatGPT. I am bit confused here. Could anyone explain it? Thanks.

[elirod]

Hi @Alan_He

Welcome to the community.

Andrew try to pointing out that diffusion models are supervised learning by heart, cause is based on neuralnetworks. Of course that there are neural networks that it is considered a unsupervised learning as well like boltzmann machine, for example.

It can relay on both category supervised and unsupervised learning, it will depend of the specific application and model.

Diffusion-Model-Assisted Supervised Learning of Generative Models for Density...

Masked Diffusion as Self-supervised Representation Learner

I hope this help

Best regards

[TMosh]

@Alan_He

or ask chatGPT

I recommend you not rely on a chat bot for answers.

[Alan_He]

Thanks all for nice feedbacks and suggestions.

[Yasar_Mehmood]

Diffusion models like generative adversarial networks are trained using supervised learning. However, the data is not annotated by human annotators. Instead, the dataset is generated/created by adding noise to an image and then treating the noisy image as a data point and its corresponding clean image (from which it was created by adding noise) as its label. During forward propagation, the network predicts the actual (without noise) image using its noisy version, and loss is the difference between the predicted and ground truth images (and this loss is fed back through backpropagation for tuning the weights). Similar methods are used in self-supervised learning, where the network learns using artificially created labels (the data used is actually unlabeled).