Probability Requirements Functor Models
Functor models learn by changes to functions rather than parameters. They seem to create opportunity for reducing energy used by AI although full formal benchmarking is required. I have filed a patent containing the IP for the functor model but am happy to work with anyone wishing to seek energy savings. All else on functor models is open research, numerous pubs exist on Zenodo.
The Probability Requirement Functor model was inspired by my experience in Object Oriented where we strove to have the code and the system closely reflect the domain and the business requirements. Far different from the neural network, this model is composed of various measures and associated probability functions corresponding to the area of interest. Conditional probabilities are represented and the model function handles the aggregation. In larger cases the model function can route to a bundle of functions. Functor SLMs for the most part can scale horizontally sometimes with CPUs too. They are composed only of functions so there is no state to manage across clusters. The functor reasoning model (FRM) is a more complex case and it does have state for CoT purposes.
2 comments
Building the model from measures and conditional probability functions that mirror the domain makes the logic auditable, which a neural net rarely offers. How do you fit or update those probability functions when the data drifts, and does that update stay inside the function-only, stateless design?
Clean idea, and the auditability is what would matter to me — a decision you can read back is worth a lot when it triggers an action rather than a suggestion. My question is where the structure comes from. If a human specifies the measures and the conditional links per domain, the energy saving is real but the coverage is narrow. If the structure has to be learned from data, the fitting problem comes back and I'd expect the update path to need state even when the runtime doesn't. Which of those two is the model you've patented?