Haskell excels at treating effects as first-class, composable entities. This talk asks about what happens when we bring that perspective into the field of machine learning. What if the implementations and specifications of training and evaluation (backpropagation, parameter initialization, regularization, stochasticity, normalization, and so forth) were not rigidly tangled spaghetti, but distinct algebraic effects tracked in row types? Could type signatures capture a model’s operational requirements directly, with handlers supplying distinct semantics for the distinct training, evaluation, and prediction stages? Treating automatic differentiation as an algebraic effect provides a foundation for reasoning about machine learning systems in the same way we reason about IO, nondeterminism, state, and aggregation. Extensible effects provide a platform for modular composition with selective elimination and experimentation. Novel operations such as hyper-parameter nudging emerge naturally as definable effects rather than ad-hoc procedures. A working toy implementation will be demonstrated for illustrative purposes. However, this phase of our research is not about achieving a polished library, instead exploring a reframing: machine learning primitives as algebraic effects, and a systematic investigation of how Haskell’s type system can give us insight into new directions and paradigms.

Exploring functional programming, type-driven development, and real-world Haskell at scale and in embedded systems.