Coherent optical neural networks drop the Mach-Zehnder mesh: a butterfly topology with one-parameter units lowers reconstruction error under tight parameter budgets
Synopsis
Coherent optical neural networks are limited by how many tunable parameters they can afford, since fabrication complexity, footprint and control overhead all grow with parameter count. The fine-grained butterfly (FGB) architecture replaces the Mach-Zehnder interferometer, which carries two tunable phase shifters, with a single phase-shifter and beam-splitter pair as the building block, cascaded in a butterfly topology so that each input-output pair has exactly one path and every path crosses the same number of units. In computational evaluations under restricted parameter counts, this architecture gives lower reconstruction error on Haar-random matrices than conventional rectangular meshes and higher test accuracy on Fashion-MNIST classification in both feedforward and recurrent networks.
Interpretation
A linear unit in a coherent optical neural network can reach strong expressive power with few tunable parameters, provided every parameter takes part in global mode mixing. Earlier coherent optical neural networks mostly relied on rectangular meshes whose building block is a Mach-Zehnder interferometer with two tunable phase shifters and two fixed beam splitters, so realizing an arbitrary unitary transformation demanded many tunable parameters and a single linear unit carried substantial overhead. In Haar-random unitary reconstruction, with parameter count normalized by the degrees of freedom of a unitary matrix, the fine-grained butterfly curves sit below those of the Mach-Zehnder mesh and its truncated variants, and the separation holds across matrix sizes.
The new architecture reduces each stage's mixing operation to a minimum: one programmable phase shifter plus one fixed 50:50 beam splitter, a unit with a single tunable parameter. A Mach-Zehnder interferometer is itself a complete two-mode interferometer and needs two parameters for one coupling, leaving fewer coupling operations under a fixed parameter budget. For 16 inputs and several butterfly blocks, with parameters initialized uniformly at random, the root-mean-square error between the analytical expected squared moduli and the computational averages is small, showing every entry has expected squared modulus 1/16.
The butterfly topology gives each input port exactly one path to every output port, with the same number of units on every path, so a statistically uniform transformation arises without specialized initialization. Rectangular meshes allow multiple paths between input-output pairs, with path counts concentrated toward central output ports, so uniform random initialization produces a banded structure; truncated meshes lose reachability outright and acquire structural zeros. In Fashion-MNIST classification, across a single-layer perceptron, a three-layer perceptron and an Elman-type recurrent network, the fine-grained butterfly architecture reaches higher test accuracy than the Mach-Zehnder mesh at a given parameter count, while the latter degrades sharply near parameter counts where reachability is lost.
Perspective
The evaluation assumes ideal devices and covers Haar-random unitary reconstruction and Fashion-MNIST classification, compared against Haar-initialized Mach-Zehnder rectangular meshes and their truncated variants; the current architecture requires input port counts that are powers of two.
No general decomposition algorithm exists for this architecture, so even when the parameter count exceeds the degrees of freedom of the unitary group it remains undetermined whether a given configuration can realize an arbitrary unitary transformation; waveguide crossings and long interstage connections add optical loss and phase error, and the practical advantage depends on the balance between reduced optical depth and routing overhead.
