nnR 0.2.0
New features
- Added
Pnm() for the general neural-network polynomials
in Definition 2.27 of Rafi, Padgett, and Nakarmi (2024).
- Added column-batch evaluation to
inst() and reusable
realizations through realize_nn().
- Added
stk_many() and nn_sum_many() for
efficient n-ary algebraic operations. Unequal-depth realization
identities are documented as ReLU-specific because their padding uses
tunnel networks.
- Generalized
MC() from one-dimensional inputs to the
full multidimensional construction in Definition 2.39, including the
valid one-sample case.
- Added
validate_nn() and repaired
draw_nn().
Correctness and API
is_nn() now verifies numeric, finite, dimensionally
connected layers.Sum(n, k) now supports every positive n
and k, as its definition requires.Prd(q, eps) now allocates the square-network error
budget using eps / (2^(q - 1) + 1).- Approximation scaling is computed in the log domain and extreme
parameters now fail with a direct floating-point-range error instead of
producing non-finite network weights.
- Composition reports incompatible end widths directly.
- Infix algebraic operators are now explicitly exported, while
implementation helpers remain internal.
- The previously auto-exported implementation objects
A,
B, A_k, C_k, and ck
are no longer public. Use Phi() or Phi_k()
instead.
- Power networks are constructed iteratively without duplicate
recursion.
- Taylor polynomials reuse a single power sequence and n-ary sum.
- Activations and realization are vectorized over complete
matrices.
- Copy, sum, identity, tunnel, norm, maximum, and maximum-convolution
constructors avoid repeated matrix growth.
- Removed package-internal
source() calls and duplicate
infix implementations.
- Added an R CMD check workflow and modernized artifact upload for the
paper workflow.