Functor Models: Auditability with Governed Learning and Repair
Functor Models have a different approach to learning as opposed to stochastic models. The functor model learns by changes to its function(s) rather than parameters. This document contains mathematical proof of distributional equivalence with a stochastic model.
Functor models are naturally one-shot learners. They are also governed. Candidates are proposed for learning and if approved are ingested into the model. A delta function for that learning unit is prepared. Then direct sum(s) of the incoming delta function and the model functions involved is computed. Any change to the model functions is logged in the version control service and a rollback method is provided if the customer specifies this. Rollback is also achievable by subtracting the deltas from the current model functions. Governance is aided by this rollback capability. The functor model can be rolled back to any prior version. Please see this paper for more information as to how the rollback capability also represents a potentially significant energy savings as well.
Figure 1 - Functor Model Logging
The example below shows how an example functor model audit trail change entry can easily be obtained via some basic integrations. Jira is shown as an example project management source but others are certainly supported.
