Want to understand the structure in the data?

[TO BE WRITTEN]

• it’s clear from CoD that there’s gotta be structure. we haven’t found it.
• Hierarchical data models (lotsa Wyart papers), DNN reverses data generating process
  • staircase fns (someone’s attempt to model the data)
• source / capacity theory: can these exponents tell you anything beyond lazy regime (maybe)?
• open questions: can you directly observe the structure in the data, independent of any learning algorithm? how do you identify HPs in data, such that you can come up with a datamodel not just for “image data” or “text data” generally, but for a particular image or text dataset? what are the right metrics for determining you got the HPs right?