When a wearable sensor sends training data to a nearby phone or access point over a wireless link, the energy cost of each sample varies widely from moment to moment. On a good channel a few kilobytes cost almost nothing. In a deep fade the radio has to raise its power sharply, by up to around 10 dB in the retransmission setting studied by Green Communication via Power-Optimized HARQ Protocols, or it has to wait. The learning value of each sample also varies. Most new samples add little once the model has seen similar data, and the value of new data rises and falls over the course of training. Today these two quantities are managed by separate systems. Machine-learning code decides which samples matter, and radio code delivers whatever bits it is handed at minimum power. Neither knows the price the other is paying.
We propose treating each sample as an item with a value (its estimated contribution to the model right now) and a price (the Joules needed to deliver it on the current channel). The device sends a sample only when its value per Joule clears a threshold. That threshold rises automatically as the model saturates. The result is a single online policy that does sample selection, transmit scheduling, power control and early stopping together. Our central claim is narrow and testable. This coupling beats the obvious "select by utility, then transmit at minimum power" pipeline only when a sample's value decays on a timescale comparable to how long the channel stays good or bad (its coherence time). We will map where that regime lies for realistic human-activity and image workloads, instead of assuming the gain is universal.
Core Claim
Two bodies of work motivate the idea. On the data side, sample value is concave in volume and saturates. Optimal Pricing of Internet of Things: A Machine Learning Approach fits service quality as $\alpha_1-\alpha_2 e^{-\alpha_3 n}$ in the number of samples $n$, and Market Model and Optimal Pricing Scheme of Big Data and Internet of Things (IoT) validates similar concave forms on UCI HAR. Both papers price data in money and never model a radio. Value is also non-stationary in time. Capturing the Temporal Dependence of Training Data Influence finds a U-shaped profile: samples matter early, matter little in a long middle basin, and matter slightly more near the end. Selecting only in the high-influence phases keeps 96% of the gains of continuous selection at one fifth of the selection cost.
On the radio side, energy per delivered bit depends strongly on channel state and on fixed circuit power. A Survey of Energy-Efficient Techniques for 5G Networks and Challenges Ahead shows that bits-per-Joule efficiency is non-monotonic in transmit power once circuit power is counted. Margin-adaptive methods such as Closed-form solution for minimizing power consumption in coordinated transmissions minimize power for a fixed rate target, but they treat every bit as equally worth delivering.
We connect the two with one per-slot energy model and one decision rule. Suppose a device sends a sample of $b$ bits in a slot of length $\tau$ over bandwidth $W$, with channel power gain $g_t$, noise density $N_0$ and circuit power $P_c$. By inverting the Shannon capacity formula, the minimum energy to deliver it is
$$E_t(b) = \tau\left(P_c + \frac{N_0 W}{g_t}\left(2^{b/(W\tau)}-1\right)\right).$$
The device keeps a small buffer. It transmits buffered sample $i$ at time $t$ when
$$\hat u_i(t) \ge \lambda_t \, E_t(b_i),$$
where $\hat u_i(t)$ is a cheap estimate of the sample's current marginal value to the model on the access point. The multiplier $\lambda_t$ is the shadow price of energy, updated online. It rises when the device overspends its energy budget, and it rises when the observed per-round gain on held-out data flattens, which is the saturation signal. When $\lambda_t$ grows large enough that nothing clears it, the device has in effect stopped early.
The update for $\lambda_t$ is a projected online gradient step on the dual of an energy-constrained utility maximization. We borrow this technique from the no-regret energy-efficiency learner of Learning to be green: robust energy efficiency maximization in dynamic MIMO-OFDM systems. That paper achieves $O(T^{-1/2})$ regret against arbitrary channel variation, and we inherit the same style of guarantee against the benchmark of a fixed price chosen in hindsight.
What Is New Relative to the Closest Work
The nearest system is Centaur, from Enhancing Efficiency in Multidevice Federated Learning through Data Selection. Its constrained devices score each sample by loss and last-layer gradient norm, then discard it, train on it locally, or send it to the access point. The authors report over 80% device energy reduction. Centaur's decision ignores the channel entirely, however. A high-value sample is sent in a deep fade at full cost, and a marginal one is discarded even when sending it would be nearly free. The authors also acknowledge that the energy spent by the access point is not modeled.
Client-level methods do incorporate wireless and energy state, but they treat each device's dataset as indivisible:
- REWAFL adjusts local iterations to uplink rate.
- Towards energy-aware federated learning on battery-powered clients weights clients by remaining battery.
- Efficient Federated Meta-Learning over Multi-Access Wireless Networks assigns resource blocks and power per client based on gradient-norm contribution.
FLrce stops the whole session when client updates start to conflict. Our proposal differs from all of these in three ways. It works per sample rather than per client. It sets the physical transmit power and timing from the value estimate. And its stopping behaviour falls out of the rising price $\lambda_t$ rather than being a separate global rule. We will also count access-point receive and compute energy in the total, which addresses Centaur's stated blind spot.
Research Program
The setting follows Centaur. Several wearable-class devices collect sensor windows or images and connect to one access point that trains the shared encoder and classifier. Each device runs a small copy of the current classifier head, refreshed each round, to compute the gradient-norm score used by Centaur. We multiply that score by a training-phase factor estimated from the access point's recent held-out improvement. This factor approximates the U-shaped influence profile without running full data-value embeddings. Those embeddings are derived for plain SGD, and the influence paper itself notes that extending them to Adam is unsolved.
Channels are block-fading Rayleigh with Gauss-Markov correlation across slots, as in the HARQ study, with an Okumura-Hata path loss model. We sweep the correlation coefficient to vary coherence time. Energy accounting includes device radiated and circuit power, device scoring compute measured on a Raspberry-Pi-class board, and access-point receive and training energy.
The analytical work has two parts. First, we derive the regret bound for the $\lambda_t$ update when values are allowed to decay with buffer age. Second, we characterize, in a stylized model with value decay rate $\delta$ and channel correlation $\rho$, the ratio of the joint policy's utility-per-Joule to that of the decoupled pipeline. We expect this ratio to approach one when $\delta$ is much smaller than the channel's decorrelation rate, because the device can then simply wait for a good channel. It should exceed one when the two rates are comparable.
Baselines
The naive reference sends every sample to the access point at fixed power, which corresponds to the offload-everything design. It uses the same model, optimizer and number of rounds as the other methods, and it defines the full-data accuracy that all targets are measured against.
The strongest learning-side competitor is Centaur's three-way loss and gradient-norm selection, run with its published thresholds. We then tune those thresholds under the same search budget we give our own price step size. The comparison uses the same device splits of UCI HAR, MotionSense, PAMAP2 and CIFAR-10. Centaur reports roughly 19% higher accuracy and roughly 58% lower training latency on average, plus over 80% device energy reduction. It does not report a number for a fading channel, and we found no reported result on Lacuna for channel-aware per-sample selection on these datasets. For a centralized accuracy ceiling on UCI HAR, HAR-Net reports 96.9%.
The strongest overall competitor is a decoupled pipeline we build ourselves. Centaur's selection feeds a buffer, and the buffer is drained by opportunistic channel-threshold scheduling with minimum-energy power control. Its channel threshold is tuned to the same energy budget. This baseline is the direct test of whether the coupling matters.
Three client-level and stopping baselines are included: REWAFL-style rate-aware local iterations, FLrce-style global early stopping, and random subsampling at the saturation volume predicted by a fitted $\alpha_1-\alpha_2 e^{-\alpha_3 n}$ curve.
Ablations isolate the key ingredients. We remove the training-phase factor, which leaves static value. We remove circuit power from $E_t$. We replace the estimated value with oracle leave-one-out value, which gives an upper bound.
Beating the decoupled pipeline means at least 20% fewer total Joules to reach 95% of full-data accuracy in the regime where our analysis predicts a gain. We must also match it within noise where the analysis predicts no gain, since a gain there would suggest a bug rather than a mechanism.
Why not just…?
Why not just select by utility and let the radio minimize power afterwards? This is the real threat, and it is exactly our decoupled baseline. If sample value barely changes while a sample sits in the buffer, waiting for a good channel costs nothing, and separating the two decisions is nearly optimal. Our claim is therefore conditional. The coherence-time sweep will show whether realistic decay rates, driven by the fast early phase and the basin transition reported in the temporal-influence paper, fall in the regime where coupling helps. If they do not, we will report that the pipeline suffices.
Why not just transmit whenever the channel is good? Channel-only scheduling captures the gains that Closed-form solution for minimizing power consumption in coordinated transmissions (30% below equal power allocation at 1 Mbps) and the HARQ work (8 to 11 dB) report for bit delivery. It still delivers redundant samples, however, which the instruction-tuning findings in What Makes Good Data for Alignment? suggest can even hurt. The oracle-value ablation separates how much of our gain comes from the channel and how much from the value estimate.
Why not just subsample to the fitted saturation point, or stop globally? A fitted concave curve says how much data to collect but not when to send it or which samples. It also ignores the non-monotonic value profile over time. FLrce's global stop saves at least 30% of compute but keeps sending redundant data until the stopping round. Both are run as baselines. If random subsampling at the fitted volume matches our method, that would show per-sample value estimates add nothing beyond knowing the saturation volume, and we would say so.
Proposed Experiments
The primary metric is total system Joules, counting both device and access point, to reach 95% of the naive baseline's accuracy. Secondary metrics are the area under the accuracy-versus-Joules curve and the worst-device accuracy. The worst-device measure checks that value-per-Joule pricing does not starve devices at the cell edge, a fairness concern the wireless survey raises.
The decisive check is a two-dimensional sweep over channel correlation and a controlled rate of value decay. The decay rate is varied by changing the learning-rate schedule and the non-stationarity of the data stream, using subject-ordered HAR streams. We plot where the joint policy's energy advantage over the decoupled pipeline exceeds 20% and compare that map with the stylized theory. The second check measures scoring overhead in Joules. The method fails on its own terms if scoring costs more than the transmissions it avoids, an open risk also flagged for online selection generally.
References
- Green Communication via Power-optimized HARQ Protocols
- Optimal Pricing of Internet of Things: A Machine Learning Approach
- Market model and optimal pricing scheme of big data and Internet of Things (IoT)
- Capturing the Temporal Dependence of Training Data Influence
- A Survey of Energy-Efficient Techniques for 5G Networks and Challenges Ahead
- Closed-form solution for minimizing power consumption in coordinated transmissions
- Learning to Be Green: Robust Energy Efficiency Maximization in Dynamic MIMO-OFDM Systems
- Enhancing Efficiency in Multidevice Federated Learning through Data Selection
- REWAFL: Residual Energy and Wireless Aware Participant Selection for Efficient Federated Learning Over Mobile Devices
- Towards energy-aware federated learning on battery-powered clients
- Efficient Federated Meta-Learning over Multi-Access Wireless Networks
- FLrce: Resource-Efficient Federated Learning with Early-Stopping Strategy
- HAR-Net: Fusing Deep Representation and Hand-crafted Features for Human Activity Recognition
- What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning