I grew up in three different states in India, shifting to Zanzibar (Africa), and now coming to the UK. My situation, on an abstract scale, rewarded understanding different cultures and context; and resistance to relying on singular domains for adaptations. It was a lucky thing for me to be learning the various disciplines that both require and are extended by Data Science and Artificial Intelligence.

Both my position across various settings, and my study culminated with a wonderful epiphany: Learning is not simply vertical after a certain depth. Knowledge across sub-domains and even across domains helps in novel work as well, albeit sparsely.

MRI for example was directly inspired from how cosmic gases are differentiated using different radiations. I noted everywhere that so much effort has gone into deriving methodologies from existing work in the same domain. Much lesser work goes into looking beyond the wall at other domains, for good reason. Since this methodology transfer is very sparse, and the value to get at human scale is less. But if an entity could encode everything about every domain, clear patterns would be much more easy to spot. But spotting methodologies as a general usecase is infeasible directly.

Thus I am working on Nodoz.

But before I get into that, it is worth sitting with the observation itself for a moment.

The MRI example is not a fluke. If you go looking, this pattern shows up everywhere. The math behind information theory was borrowed wholesale into genetics. Neural networks were loosely inspired by how we thought neurons fired. Shazam’s audio fingerprinting works on principles used in radar signal processing. None of these were obvious at the time. They required someone to be standing at the intersection of two fields and have the presence of mind to notice the structural resemblance.

That is the thing about cross-domain transfer: it is not about surface similarity. The connection between cosmic gas differentiation and soft tissue imaging is not immediately obvious unless you abstract both down to the same skeleton — a problem of distinguishing between materials using differential response to varying frequencies. At that level of abstraction, they are the same problem. And once you see that, the methodology follows.

The reason this does not happen more often is not that researchers are incurious. It is that maintaining the depth required to work at the frontier of your own field, while also maintaining enough breadth to notice structural patterns in unrelated fields, is genuinely hard. Most people cannot do both. The ones who can are rare, and they tend to produce disproportionate work.

What I kept wondering was whether this had to remain a property of rare individuals, or whether it could become a property of a system.

That question is what I have been sitting with for the past year or so. The research end of it is concerned with how you even represent a methodology in a way that is domain-agnostic, so that the structural match between an astrophysics paper and a medical imaging paper becomes computable rather than intuitive. The commercial end of it is more grounded: given a dataset, what methodology should a researcher try, drawing from the full breadth of scientific literature rather than just their own field?

Both questions are hard. I think they are worth working on. More on Nodoz separately [→ link].