0.4.0

New Features

• Add an Ising layer for performing quantum annealing-based neural
  network computations. Its backpropagation is computed, or
  approximated, using sample covariance.

• Add statistic base class for untrainable statistics, e.g., sufficient
  statistics of Ising models.

• Add a static affine module. It is static in the sense that there are
  no trainable parameters.

• Allow hidden units to be connected during GraphRestrictedBoltzmannMachine
  construction.

• Add BipartiteGibbsSampler for performing Gibbs sampling of bipartite
  graph-restricted Boltzmann Machines.

• Add conditional sampling functionality for the BlockSampler.

Upgrade Notes

• Add .clone() to the return of BlockSampler.sample to prevent
  unintended in-place modification of the sampler's internal state due
  to returning a reference to the underlying tensor.

• Add conditional sampling functionality for the DimodSampler.

• Initialize GraphRestrictedBoltzmannMachine weights using Gaussian
  random variables with graph-connectivity-dependent standard
  deviations. For an edge $(u, v)$, the default standard deviation is
  $2.5 / (\deg(u)\deg(v))^{1/4}$. The weight-initialization strategy is
  grounded in Hinton's practical guide for RBM training, which
  recommends sampling weights from a Gaussian distribution with mean 0
  and standard deviation 0.01 (for zero-one-valued RBMs). The
  connectivity scaling keeps the energy functional extensive on sparse
  graphs, while the temperature factor initializes the GRBM deep in a
  paramagnetic regime for QPU-backed sampling, consistent with the
  Sherrington-Kirkpatrick model.

Bug Fixes

• The Gaussian kernel incorrectly computed pairwise distances with l2
  norm without squaring it. This fixes the bug by squaring the l2 norm.

• GraphRestrictedBoltzmannMachines should not be allowed to be defined
  with self-loops. Presence of self-loops is now checked at construction
  time.

• Raise a ValueError when GraphRestrictedBoltzmannMachine.set_quadratic
  receives an edge that is not in the model. Previously, an unknown edge
  overwrote every quadratic bias.