I am really interested topics like disentanglement and generative models and recently red this paper: https://arxiv.org/abs/2205.10268
I was wondering if replacing affine transformations by B-cos transformations (which is enforcing alignment between the weights and the input) could be benefitial in a generative setting. Furthermore, could it not be helpful to use these new layers to find disentangled latent factors? I was looking through all the citations but nobody uses B-cos networks in a generative framework. Maybe someone has more intuition about this topic or heard about related papers that already tried out similar stuff which he can share.
The first (and only problem) I see is, that the weight-input alignment is achieved by maximizing class logits in a classification problem. In a generative setting we will probably have a completly different loss function/optimization problem.

I would be happy to open a discussion and hear and talk about your/my ideas and opinions!