A mixture-of-experts model is a neural network split into several sub-networks ("experts") plus a small model, the router, that decides which experts process each input. When the experts sit on different wireless edge devices, every routing decision is also a transmission decision. If a feature is sent over a fading link, it arrives corrupted. The closest prior work, Mixture-of-Experts for Distributed Edge Computing with Channel-Aware Gating Function, feeds the router the current signal-to-noise ratio of each expert's link. That router learns to send features to a less suitable expert when the best expert's link is noisy. It treats the radio channel as a fixed obstacle to be routed around.
Reconfigurable intelligent surfaces change that assumption. These are walls or panels covered with hundreds of cheap, nearly passive elements whose reflection phase can be switched electronically, which lets the environment itself be steered toward a receiver (Reconfigurable Intelligent Surfaces: Principles and Opportunities; Smart radio environments empowered by reconfigurable AI meta-surfaces). This note proposes treating the surface configuration as a second routing decision, one level above the per-input expert choice. For each block of inputs, the system jointly picks a surface configuration and the expert assignments, so that the channel is opened toward the experts the data actually needs. The case where this is not trivial is the interesting one. A single surface configuration is shared by every transmission in a coherence block, the short interval over which the channel stays roughly constant. Inputs in the same block that want different experts therefore compete for the same physical resource.
Core Claim
Conditional computation (running only part of a network per input) can be extended so that part of the "computation graph" is the propagation medium. The routing problem then has two coupled levels: a per-input expert choice and a per-block environment choice. The claim is that solving them jointly beats both existing alternatives when several experts must be reached within one block. The first alternative, from Wang et al., adapts the router to whatever channel exists. The second is the obvious engineering pipeline: let the router pick experts from the data alone, then point the surface at the chosen experts. Both fail in the same regime. The surface cannot favor every link at once, so the best configuration depends on which experts the whole block needs and how much each input loses by being sent elsewhere. Neither pipeline sees both quantities.
Concretely, let $\mathcal{C}$ be a small codebook of quantized surface configurations, with 1- or 2-bit phases as in practical hardware (Novel 1-bit Hybrid Reconfigurable Intelligent Surface). Let $s(\theta)$ be the vector of per-expert link SNRs measured under configuration $\theta$. Let $G(z, s)$ be a channel-aware router that maps a feature $z$ and SNR vector $s$ to expert weights. Let $\hat\ell$ be a learned predictor of task loss after noisy transmission. For a block of features $z_1,\dots,z_B$, the controller picks
$$\theta^\star = \arg\min_{\theta \in \mathcal{C}} \sum_{i=1}^{B} \hat\ell\big(z_i,\, G(z_i, s(\theta))\big),$$
and then routes each $z_i$ with $G(z_i, s(\theta^\star))$. Training relaxes the argmin with a Gumbel-softmax over $\mathcal{C}$, so the router, the loss predictor, and a configuration-scoring head learn together. Restricting the choice to a codebook is deliberate. Estimating the full cascaded channel through a passive surface scales with element count, so measuring $|\mathcal{C}|$ beam-training pilots is the realistic interface.
Research Program
We reuse Wang et al.'s two-stage recipe as the method backbone. Experts and backbone are trained on clean features and then frozen, and only the routing side is trained under simulated channels, with their balance regularizer to stop the router collapsing onto historically good links. On top of that we add three things: the configuration codebook, the block-level selection above, and a channel simulator. The simulator places a dispatcher, four to eight expert nodes, and a surface of 64 to 256 elements in an indoor geometry, with the direct path partly blocked and Rician or Rayleigh fading. The coherence block length $B$ and the diversity of the data within a block are explicit knobs. Low diversity means most inputs want one expert. High diversity means they want several. The theory predicts the joint method's advantage should grow with how many distinct experts a block needs and shrink to zero when it needs one.
Two physical constraints from the surface literature shape the design and give falsifiable predictions. The first is the "double-fading" path loss: reflected power is the product of two link losses. Integrating Over-the-Air Federated Learning and Non-Orthogonal Multiple Access: What Role can RIS Play? finds a surface halfway between endpoints is the least useful placement. We therefore expect our gains to collapse at midpoint placement and will sweep surface position to check. The second is phase quantization. The same paper reports a further gain going from 1-bit to 2-bit phases, so we evaluate both. That paper also supplies the closest precedent for using a surface to arbitrate between competing uses of one channel, separating decodable users from aggregated ones. It optimizes a rate objective with iterative convex solvers and has no learned router, which is the piece we add.
The cellular-heterogeneity direction enters only as an evaluation lesson, not as a mechanism. Bulk measurements average away the subpopulations that matter, and deconvolution benchmarks show that aggregate scores hide systematic failure on specific groups (Performance of tumour microenvironment deconvolution methods in breast cancer). The analogous risk here is that a mean-accuracy gain hides worse service for blocks dominated by experts in shadowed positions. We therefore report accuracy per expert location and per class, not only the average. We do not claim any methodological transfer from biology.
Baselines
The naive reference is a standard sparse mixture-of-experts router of the GShard type (GShard). It routes on data features only and runs with a random surface configuration per block. It is the floor that shows how much channel distortion costs.
The strongest existing method is Wang et al.'s channel-aware router, run with no surface control. On their benchmark it lifts ResNet-18 on CIFAR-10 in analog transmission from 84.7% to 91.1%, and ViT on CIFAR-100 in digital 16-QAM transmission from 58.2% to 62.0% (Mixture-of-Experts for Distributed Edge Computing with Channel-Aware Gating Function). We first reproduce these numbers in our simulator with the surface fixed, to confirm the comparison is calibrated. We then give the same router a random configuration and a fixed best-on-average configuration.
There is no reported result found on Lacuna for any mixture-of-experts system that controls a reconfigurable surface. The closest surface-side number is Ni et al.'s throughput gain over random configurations, "up to 22.3%". That is a rate, not an accuracy, so it only tells us surface control can matter at the link level.
Two ablations isolate the key ingredient, which is coupling the configuration choice to the router's per-input losses:
- Channel-only controller. The configuration maximizes the sum or minimum SNR across experts, and Wang's router then adapts to it.
- Demand-driven controller. The data-only router picks experts first, and the configuration is then chosen to maximize SNR toward exactly those experts, weighted by how many inputs chose each.
An oracle upper bound uses exhaustive joint search with true continuous-phase CSI.
All methods share:
- the same frozen experts;
- the same codebook and pilot budget, with $|\mathcal{C}|$ measurements per block;
- the same router architecture and training steps;
- the same hyperparameter search budget;
- the same transmit power and seeds.
Beating the strongest competitor means higher accuracy than the better of the two decoupled controllers by more than three seed standard deviations at matched SNR, in the high-diversity, low-SNR regime. An equivalent criterion is reaching their accuracy with measurably less transmit power. Matching them only when blocks need one expert is predicted and does not count.
Why not just…?
Why not just pick experts from the data and then point the surface at them? This demand-driven pipeline is the real threat. When a block needs one expert it is optimal, and our method should only tie it. It breaks when inputs in a block disagree. It weights experts by vote count rather than by how much each input loses if rerouted, so one hard input that only expert 3 can handle gets outvoted by many easy inputs that any expert could serve. The diversity sweep settles this directly. If the demand-driven controller matches the joint one at all diversity levels, the honest conclusion is that the simple pipeline suffices, and we will report it that way.
Why not just raise transmit power or add more experts? Wang et al. note that more distributed experts help by offering alternative paths under noise. Power, though, is bounded by regulatory and thermal limits, and extra experts add hardware. The surface is nearly passive and has no amplifiers (Smart radio environments empowered by reconfigurable AI meta-surfaces). We report the transmit-power increase, and separately the number of extra experts, that the strongest baseline needs to match our accuracy. That converts the comparison into decibels and devices saved. If this saving is small, the case for surface control is weak.
Why not just optimize the surface for link quality and let Wang's router adapt? This channel-only controller is what the surface literature already does, optimizing rate as in Ni et al. Its blind spot is that it spends reflected power on experts no input in the block needs. The ablation measures exactly this. A useful diagnostic is the fraction of blocks where the SNR-optimal and loss-optimal configurations differ. If that fraction is near zero, the joint objective adds nothing.
Why not just use digital transmission with error correction? Coding removes small errors but turns deep fades into lost packets, and Wang et al. show channel-aware routing still helps in their digital 16-QAM mode. We evaluate both analog and digital modes so the result does not depend on one transmission scheme.
Proposed Experiments
The main experiment reproduces Wang et al.'s CIFAR-10 / ResNet-18 analog and CIFAR-100 / ViT digital settings inside the surface-equipped simulator. It sweeps three things, with the 1-bit and 2-bit codebooks reported for each:
- SNR, from 0 to 20 dB;
- block diversity, measured as the number of distinct preferred experts per block;
- surface placement, near the dispatcher, near the experts, and at the midpoint.
Metrics:
- top-1 accuracy, as mean and per-location;
- transmit power needed to hit a target accuracy;
- pilot overhead per block;
- controller compute time relative to block duration.
The joint argmin over a small codebook must be cheap enough to finish inside a coherence block, and we time it rather than assume it.
Three decisive checks decide whether the idea holds:
- Diversity. The advantage over the demand-driven controller must increase with block diversity and vanish at diversity one. A flat or absent trend falsifies the core claim.
- Placement. Gains must shrink at midpoint placement, as the double-fading argument predicts. Gains that survive there would suggest the simulator, not the method, is responsible.
- Distribution of gains. Per-location accuracy must not reveal that the mean gain is bought by degrading shadowed experts. If it does, we report the trade-off rather than the average.
References
- Mixture-of-Experts for Distributed Edge Computing with Channel-Aware Gating Function
- Reconfigurable Intelligent Surfaces: Principles and Opportunities
- Smart radio environments empowered by reconfigurable AI meta-surfaces: an idea whose time has come
- Novel 1-bit Hybrid Reconfigurable Intelligent Surface
- Integrating Over-the-Air Federated Learning and Non-Orthogonal Multiple Access: What Role Can RIS Play?
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
- GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding