Decentralized learning lets devices train a shared model by repeatedly averaging with peers instead of reporting to a server. To save bandwidth, a device can stay silent until its model has drifted enough from the copy its peers already hold. This "event-triggered" rule, where a message is sent only when a tracked error crosses a threshold, has worked well in settings where every peer is one radio hop away. Real wireless edge networks are not like that. Two devices that are logical neighbors in the averaging scheme may be several physical hops apart, so one "message" is really a chain of retransmissions through relays.
This note proposes a simple change: price each logical connection by how many radio transmissions it actually costs, and let a device transmit over a long, relayed connection only when its update is worth that many transmissions. With this pricing, a few relayed long-range links on top of ordinary neighbor-to-neighbor averaging should give more accuracy per radio transmission than two alternatives. The first is pure one-hop event-triggered averaging. The second is a dense logical overlay that ignores hop cost. The benefit should be largest for knowledge held by devices on the edge of the network. The note also tests whether a censoring-aware version of the "effective number of neighbors" can predict where the best operating point lies before training.
Core Claim
Existing event-triggered decentralized learners treat every message as costing the same, because every neighbor is physically adjacent. EF-HC (extended version) lets node $i$ broadcast only when its model has moved more than $r\rho_i\gamma^{(k)}$ from its last broadcast. Here $\rho_i$ reflects the node's local bandwidth, and the paper lists topology-aware threshold design as future work. DETSGRAD uses one threshold scaled by the learning rate, tested on rings. Liu et al. use summable threshold sequences and a global clock. None of the three models relaying, so none can tell a cheap adjacent link from an expensive five-hop one.
Our claim is that the threshold should belong to the edge rather than the node, and should grow with the physical path length. For a logical edge $(i,j)$ realized over $h_{ij}$ radio hops, node $i$ sends its current model $w_i^{(k)}$ to $j$ at step $k$ only if
$$\|w_i^{(k)} - \hat w_{i\to j}^{(k)}\| > r\,\rho_i\,h_{ij}^{\alpha}\,\gamma^{(k)}.$$
In this rule, $\hat w_{i\to j}^{(k)}$ is the copy of $i$'s model that $j$ currently holds, $r$ and $\gamma^{(k)}$ are EF-HC's global scale and decaying sequence, $\rho_i$ is EF-HC's bandwidth factor, and $\alpha \ge 0$ controls how strongly hop count raises the bar. Setting $\alpha=0$ recovers EF-HC applied to a logical overlay. The receiver keeps mixing its cached copy while the sender is silent, so silence on a long edge costs staleness, not lost information.
We pair this with one relay-side rule. If a relay's queue already holds an older model from the same source to the same destination, the newer packet replaces it. Relays never forward superseded models.
Why should long edges be worth paying for at all? Two findings motivate this. Sakarvadia et al. show that knowledge held only at peripheral nodes gets "washed out" as it diffuses hop by hop through averaging. The DecAvg topology study similarly finds that averaging at well-connected hubs dilutes rare updates, and that modular clusters trap local knowledge. A relayed long edge carries a model across the network without diluting it at every intermediate node. It is the routing-layer counterpart of Sakarvadia et al.'s centrality reweighting, and unlike that method it does not need global betweenness scores. The hop-priced trigger is what keeps these edges from overwhelming the radio channel.
Research Program
The work has three parts.
Simulator. We build a discrete-event wireless simulator. Devices are placed as a random geometric graph, and each transmission occupies a shared medium with contention, so that relay queues, back-pressure, and delay emerge naturally rather than being assumed away. Huang et al.'s 802.11 ad hoc scalability study motivates this care. They report that standard beacon synchronization breaks down at around 80–100 stations. That is exactly why we build on EF-HC's asynchronous event structure rather than on Liu et al.'s globally clocked scheme.
Logical graph. Every device averages with its physical neighbors. In addition, each device maintains $L$ long-range logical edges, routed by shortest path. We compare random choices of the far endpoint with choices that spread endpoints across detected communities. Community detection uses only local information, such as label-propagation communities or hop-distance sketches. This avoids Sakarvadia et al.'s reliance on global betweenness.
Predictive measure. Vogels et al. (Beyond spectral gap) show that the spectral gap mispredicts decentralized training. They propose the effective number of neighbors $n_W$, which measures how much a random walk on the mixing matrix $W$ reduces variance. We extend this in two steps. First, we compute $n_{\bar W}$ on the time-averaged realized mixing matrix $\bar W$, which already includes censoring and relay delay. Second, we divide by the realized physical transmissions per step. We then test whether this cost-normalized quantity, measured in a short pilot run, ranks configurations of $(L, \alpha, r)$ in the same order as final accuracy per transmission.
Vogels et al.'s analysis assumes IID data. We therefore also report how the ranking degrades under non-IID splits, which is the open problem they leave. We do not promise a convergence proof for non-convex models. Our target is an empirical predictor plus a convex-case sanity check that inherits EF-HC's analysis with edge-wise thresholds.
Baselines
The naive reference is synchronous decentralized averaging over the physical one-hop graph, with every node sending every step. It fixes the accuracy ceiling for that graph and the communication cost we are trying to cut.
The strongest existing method is EF-HC run on the physical one-hop graph, with its bandwidth-scaled node thresholds, since it is the only published method in the file that combines event triggering, heterogeneous devices, and decentralized FL. The file reports that EF-HC achieved higher accuracy per unit transmission time than global-threshold and zero-threshold baselines on Fashion-MNIST, FEMNIST, and CIFAR-10. However, no reported result found on Lacuna for EF-HC's exact accuracy or savings figures. The closest quantified anchor is DETSGRAD on a 10-agent MNIST ring. It reports a 74.2% communication reduction under IID data and 66.8% under single-class non-IID data, at about 98.4% accuracy versus 98.6% for a centralized model (DETSGRAD). We rerun DETSGRAD on our rings to check that our simulator reproduces these numbers before trusting any comparison.
Two further baselines complete the picture. The first is the dense-overlay baseline: EF-HC applied to the same logical graph as ours, with $\alpha=0$. The second is Topology-Aware Aggregation (Sakarvadia et al.) on the physical graph, given oracle global betweenness. That makes it an upper bound on reweighting without relays. On its own benchmarks it lifts out-of-distribution accuracy to 42% from a 20% baseline.
All methods share the same data splits, model, placements, and simulator seeds. Each gets the same budget of 30 hyperparameter trials, tuned for accuracy at a fixed transmission budget. Transmissions are counted as bits multiplied by physical hops.
"Beating" the strongest baseline means one of two things, at a matched physical-transmission budget. Either the proposal reaches higher worst-group accuracy, meaning accuracy on classes held only by peripheral devices, with average accuracy within 0.5 points. Or it reaches the same accuracy with at least 25% fewer transmissions, across three placements.
Ablations isolate each ingredient. Setting $\alpha=0$ removes hop pricing. Setting $L=0$ removes long edges. Disabling the relay-replacement rule removes the freshness mechanism. A final ablation spends the same transmission budget on extra physical-graph averaging rounds, to test whether long edges matter at all.
Why not just…?
Why not just gossip with physical neighbors and never relay? Vogels et al. show that local averaging alone trains well on rings and tori, so this is the most serious threat. Their evidence, however, is for IID data. Under non-IID splits, the DecAvg study reports peripheral-only digit classes "often hovering around a baseline accuracy of 0.5." Sun et al. also find that decentralized training on a ring falls nearly 20 points of CIFAR-10 accuracy below centralized training with 30% client participation. The $L=0$ ablation and the matched-budget extra-rounds ablation settle this directly. If one-hop EF-HC matches us on worst-group accuracy per transmission, we will report that relaying is not worth its cost. That would be a useful negative result.
Why not just reweight neighbors by centrality instead of adding routes? Sakarvadia et al. do exactly this and get large out-of-distribution gains. Their method needs global betweenness, and it still moves peripheral knowledge only through repeated diluting averages. We give it the oracle and compare. If it matches us, then the benefit comes from weighting rather than routing, and the hop-priced trigger should be tested on top of it instead.
Why not just compress every message and skip censoring? Compression, as surveyed in the federated overhead literature, shrinks each message but still pays $h$ transmissions every step on every long edge. Censoring removes whole messages. The two approaches are complementary. We run top-k compression at matched bits-times-hops, alone and combined with our trigger, to measure whether the gains stack.
Why not just raise EF-HC's global threshold $r$? A uniform increase silences cheap adjacent links just as much as expensive relayed ones. The $\alpha=0$ ablation, with $r$ tuned under the same 30-trial budget, tests whether edge-wise pricing matters beyond simply sending less overall.
Proposed Experiments
Benchmarks. We use three settings.
- The DecAvg MNIST split, in which 10% of nodes hold digits 5–9. These nodes are placed on the geometric periphery.
- Fashion-MNIST non-IID with LeNet-5 at 10–30 nodes, following EF-HC.
- CIFAR-10 with ResNet-18 at 100 nodes under a Dirichlet label split with $\beta=0.1$, following Sun et al.
Metrics. The primary metric is accuracy versus cumulative physical transmissions. We also report worst-group accuracy and wall-clock time under contention.
Decisive checks. Four results would decide the question.
- On the peripheral-digit split, does the proposal lift worst-group accuracy clearly above the roughly 0.5 plateau at a transmission budget no larger than one-hop EF-HC's?
- Does the gap between the proposal and $\alpha=0$ grow with network diameter? Hop pricing should matter more when paths are longer.
- Does the cost-normalized $n_{\bar W}$ from a 5% pilot run rank-correlate (Spearman correlation above 0.7) with final accuracy per transmission across the $(L,\alpha,r)$ grid, under both IID and non-IID data?
- Does the relay-replacement rule reduce the staleness of cached models without reducing accuracy?
A clear failure on check 1 against one-hop EF-HC would falsify the core claim.
References
- Decentralized Event-Triggered Federated Learning with Heterogeneous Communication Thresholds
- Event-Triggered Decentralized Federated Learning over Resource-Constrained Edge Devices
- Distributed Stochastic Gradient Descent with Event-Triggered Communication
- Resource-Aware Exact Decentralized Optimization Using Event-Triggered Broadcasting
- Topology-Aware Knowledge Propagation in Decentralized Learning
- Impact of network topology on the performance of Decentralized Federated Learning
- On the scalability of IEEE 802.11 ad hoc networks
- Beyond spectral gap: the role of the topology in decentralized learning
- Which mode is better for federated learning? Centralized or Decentralized