Choosing Which Buffered Sample to Send When Noise Comes in Bursts

A sensor that feeds a remote machine-learning model over a slow link has to decide, at every transmission opportunity, which of its recently recorded samples to send and how long a stretch of history to include. The standard rule of thumb is to send the freshest data. Recent work on remote inference has shown that this rule is often wrong. Measured by the eventual prediction error, an older sample, or a longer and therefore later-arriving window, can beat the newest one when the monitored process has periodic or long-memory structure. Every scheduler built on that insight, however, assumes the recorded samples are equally clean. Real sensors sit in environments where interference arrives in bursts: a siren, a door slam, a passing truck, a wave of babble. In those environments the newest sample may be the one ruined by a transient. A slightly older sample recorded during a quiet moment may carry far more usable information.

This note proposes a scheduler that looks at the measured noise level of each buffered sample, not only at its age. It treats prediction error as a joint function of how old the sent window is, how long it is, and how corrupted it is. It picks the buffer entry and window length that minimize that error. The question we want to settle is concrete: in which regimes does this joint decision beat the two obvious simplifications? The first simplification is "send the freshest sample that is clean enough," which ignores the age structure. The second is "send by age and let a noise-robust receiver cope," which ignores the noise.

Core Claim

Age of Information (AoI) is the time elapsed since the newest delivered sample was generated. The selection-from-buffer line of work showed that inference error need not grow with AoI. How Does Data Freshness Affect Real-time Supervised Learning? introduced the model in which the transmitter keeps a sliding buffer and picks the historical sample minimizing expected error. It solved the single-source case with Gittins-index thresholds and the multi-source case with Whittle indices. On the Monotonicity of Information Aging explained the non-monotonicity through ε-Markov chains, a measure of how far a process departs from "the present summarizes the past." Learning and Communications Co-Design for Remote Inference Systems and Goal-Oriented Status Updating for Real-Time Remote Inference Over Networks With Two-Way Delay added the choice of feature length, meaning the number of consecutive samples per packet. Longer windows give the predictor more context but take longer to transmit.

These papers supply our method. Their decision states, however, contain only age, delay, and buffer position. Their evaluations use stationary sources such as AR processes, Jakes fading, CartPole, and BAIR. They are blind to content by construction.

Our claim is that under bursty, non-stationary observation noise, the loss the scheduler should minimize takes the form

$$ \min_{\pi}\ \limsup_{T\to\infty}\frac{1}{T}\,\mathbb{E}_\pi\!\left[\sum_{t=1}^{T} L\big(\Delta(t),\, l(t),\, \mathbf{q}(t)\big)\right], $$

where $\pi$ is the transmission policy and $\Delta(t)$ is the age of the delivered window at time $t$. The window length is $l(t)$. The vector $\mathbf{q}(t)$ holds the cheaply estimated per-frame signal-to-noise ratios (SNR) inside that window. $L$ is the expected loss of the trained receiver given those three quantities.

Adding $\mathbf{q}$ changes the structure of the optimal policy in two predictable ways. First, the freshest-is-best threshold becomes a family of SNR-dependent age thresholds: the policy accepts an older sample only if it is cleaner by a margin that grows with the age gap. Second, the optimal window length shrinks when a burst is present in the buffer. This is the context-versus-contamination tension the research file identifies. Longer windows act as a "Markovian bridge" in clean data, per The Age of Correlated Features in Supervised Learning based Forecasting. The same longer windows are more likely to span a transient.

We make one sharp, falsifiable prediction. When the monitored process is close to Markov ($\epsilon\approx 0$), the simple rule "freshest sample above an SNR gate" should be near-optimal and our scheduler should add little. When the process is periodic ($\epsilon$ large), the cleanest sample and the best-phased sample often differ, and only the joint policy should capture the gap. If the gain does not concentrate where $\epsilon$ is large, the claim is wrong.

Research Program

The work has three parts.

The first part is an estimated error surface. For each task we train the receiver once with jointly appended AoI and SNR inputs, extending the "joint training with AoI" idea of Shisher et al. (2021), which matched separately trained models at a fraction of the cost. We then estimate $L(\Delta, l, \mathbf{q})$ on held-out data, with $\mathbf{q}$ compressed to a few quantized summaries such as the minimum and mean frame SNR of the window. The SNR estimate must run on a microcontroller. We therefore use frame energy plus a voiced/unvoiced check of the kind used for noise bootstrapping in A Two-step NMF Based Algorithm for Single Channel Speech Separation, and the frame-level instantaneous SNR that Yu et al. (2008) showed is informative for tracking non-stationary conditions. We do not run a full enhancement network at the sensor. The file documents that dual-branch and state-space enhancers are too heavy for edge hardware.

The second part is policy structure and computation. We model the noise as a Markov-modulated burst process with quiet and burst states, mirroring the piece-wise stationary modeling of Age Optimum Sampling in Non-Stationary Environment. That paper applied this modeling to network delay; we apply it to the observation itself. We extend the semi-Markov decision formulation of Arı et al. (2026) by adding the quantized SNR of each buffer entry to the state. With the small buffers those papers found sufficient ($B=5$ in Shisher et al. (2023)) and three or four SNR levels, the enlarged state is still solvable by value iteration offline.

We also seek a closed-form statement for the Markov-source case: the optimal choice is the freshest entry whose SNR exceeds a threshold that increases with its age. This gives a lookup-table policy cheap enough for the sensor, which addresses the file's complexity concern. For the online setting, where noise statistics drift, we borrow the Kolmogorov-Smirnov change-point restart of Zhang et al. (2023). We apply the test to the SNR stream rather than to the delay stream, so the burst-process parameters are re-estimated after a regime change.

The third part is evaluation on two testbeds. They are described under the experiments below.

Baselines

The do-nothing reference is the generate-at-will, freshest-sample policy with fixed length and a receiver trained without age or SNR inputs. This is the operating point the AoI-for-inference papers report large improvements over. For example, Shisher & Sun (2022) report a 3x error reduction over generate-at-will on BAIR. The reference tells us how much of the gain is already available without any noise awareness.

The strongest existing method is the content-blind goal-oriented policy of Arı et al. (2026). It selects buffer entry, variable packet length, and waiting time from age and delay state. On CartPole-v1 that paper reports up to a 15.7% error reduction over constant packet length, and errors as low as one-sixth of those of unit-length age-based scheduling. This is the right competitor because it already exploits non-monotonic aging and the length trade-off, everything except the noise. We give it the same estimated error surface marginalized over SNR, the same receiver network, the same buffer size, the same delay model, and the same value-iteration budget. Beating it means a lower time-averaged inference error with non-overlapping confidence intervals across seeds and noise draws, on the noisy versions of both testbeds. It should also stay within noise of it on the clean versions, where our method should reduce to it. For multi-source runs we also include the Whittle index policy of Shisher & Sun (2022), which they report gives a 2x gain over Maximum Age First on BAIR.

Three ablations isolate the key ingredient. The first is an SNR-gated freshest policy: send the newest entry above a tuned SNR threshold, with fixed length, the tuning grid matched in size to our policy's tuning. The second is our policy with length frozen, which isolates the window-shrinking effect. The third is "always send the cleanest entry," which ignores age altogether. If the joint policy cannot beat the gated-freshest ablation on periodic sources, the core claim fails.

Why not just…?

Why not just denoise before inference and then schedule by age as usual? This is the most serious alternative. Strong enhancers do exist: SN-Net, described in Interactive Speech and Noise Modeling for Speech Enhancement, reports 19.52 dB SDR on the DNS Challenge, and EDNet reports a real-time factor below 0.04. Running them at the sensor is exactly what the file flags as infeasible on microcontrollers. Running them at the receiver cannot restore content that a loud burst has masked. It also does nothing about the scheduler wasting a slow channel on a ruined window. We test this directly with an arm that places a receiver-side enhancer before the predictor under the content-blind schedule. If that arm matches our scheduler, the scheduling contribution is not worth its complexity, and we will say so.

Why not just gate on SNR and otherwise send the freshest sample? This is the cheap heuristic, and on Markov-like sources we expect it to be nearly as good. That is our own prediction. Its weakness should appear only when phase matters: a quiet sample a quarter-period old can be worse than a mildly noisy sample one full period old. The sweep over the periodic-source strength, described below, settles whether this weakness is real.

Why not just make the receiver robust, for example with test-time adaptation, and leave the scheduler alone? Adapting only the normalization statistics, as AdaKWS does, raises keyword accuracy under babble from 26.44% to 49.56%. The same paper, however, reports that the best batch size is 128. Large batches smooth over exactly the short bursts we care about, so receiver adaptation tracks slow drift rather than transients. We include an AdaKWS-adapted receiver under both the content-blind schedule and our schedule. The two should be complementary; if they are redundant, that will show.

Proposed Experiments

The first testbed is CartPole-v1 remote state prediction, the benchmark shared by Arı et al. (2026) and Shisher et al. (2023). We corrupt the sensor observations with additive noise whose variance follows real burst envelopes. These envelopes are extracted from DEMAND recordings and from ESC-50 environmental sounds (dog barks, sirens, door slams). We add a periodic driving force with tunable strength to sweep the source from near-Markov to strongly periodic. Its clean version reproduces the existing setting, so we can check that our implementation of the Arı et al. baseline recovers its reported behavior before adding noise.

The second testbed is streaming keyword spotting on Google Speech Commands with MS-SNSD and ESC-50 noise inserted as bursts rather than stationary beds. The sensor buffers overlapping one-second windows and sends them over a link with random two-way delay. The receiver must classify the most recent utterance before a deadline. Here the metric is accuracy at deadline. The AdaKWS result above is the closest reported figure for this noise setting. No reported result was found on Lacuna for any scheduling policy on Speech Commands, so all scheduling comparisons there are new measurements.

Metrics are time-averaged inference error (state MSE for CartPole, deadline accuracy for keyword spotting). We also report the chosen window length as a function of burst presence, and scheduler compute per decision on a microcontroller-class budget.

The decisive checks are three. The first is whether the improvement of the joint policy over the gated-freshest ablation grows with the periodic-forcing strength, and stays near zero when forcing is off. The second is whether the chosen window length drops measurably when a burst is in the buffer; if it does not, the length component of the claim is empty. The third is whether the lookup-table threshold policy stays within a few percent of the full value-iteration policy, which determines whether the method is deployable on the sensor at all. A final online run switches the noise corpus mid-stream to test whether the SNR-based change-point restart recovers faster than a policy with fixed burst statistics.