Workgroups

Instead of using a fixed network structure, how could a Boltzmann machine grow by adding latent variables in response to the difference between "wake" statistics (where the observed variables are clamped by observation) and "sleep" statistics (where even the observed variables are sampled according to the network's energy levels)?

This will be a brainstorming session. If you like Ising models, Gibbs sampling, and generative models, come brainstorm with us.