A spiking neural network processes information as sparse binary pulses ("spikes") over time, so its computation costs almost nothing when nothing is happening. A wake-up receiver is a tiny always-on radio, drawing microwatts or less, that switches on a sleeping processor only when a transmitter asks for it. Together they promise a sensing node that spends essentially zero power between events. The best existing design of this kind, Neuromorphic Split Computing With Wake-Up Radios, tunes its many detection and decision thresholds in two steps. It first searches in a software simulator of the wireless channel (a "digital twin"). It then certifies the chosen thresholds on the real channel with a statistical test, which guarantees an error budget even if the simulator is badly wrong. That guarantee hides a cost. If the twin is wrong, the system does not fail. It becomes conservative: it keeps the radio and processor awake longer and returns larger, less useful sets of candidate labels. Under risk-controlled calibration, the gap between simulation and silicon is paid in energy, not in errors.
We propose to measure that cost and remove most of it. The twin in prior work treats wake-up as instantaneous and lossless. Real hardware needs a stochastic start-up delay, suffers a supply-voltage dip at power-on, and loses the neurons' internal state whenever the processor is power-gated. On genuinely temporal data, such as spoken digits encoded as spikes, the first milliseconds of a spike stream carry information, and a processor that is still booting throws them away. We will build a transient-aware twin calibrated from a few hardware measurements, train spiking networks that expect this truncation and reset, and show that the certified operating point then consumes measurably less energy at the same guaranteed error rate. The effect should be large on temporal benchmarks and close to zero on N-MNIST, and that contrast doubles as a built-in falsification test.
Core Claim
When thresholds are certified with Learn-Then-Test (a procedure that tests candidate settings in a fixed order and only accepts those whose measured risk passes a concentration bound), the twin decides which candidates are tested. The hardware test decides only whether they pass. Write $\lambda$ for the vector of thresholds (event detector, wake-up trigger, inference confidence), $R(\lambda)$ for the true miscoverage on hardware, and $E(\lambda)$ for the expected energy per sensing window. The deployed setting is
$$\hat\lambda = \arg\min_{\lambda \in \Lambda_{\text{twin}}} E(\lambda) \quad \text{subject to LTT certifying } R(\lambda) \le \alpha \text{ with probability } 1-\delta,$$
where $\Lambda_{\text{twin}}$ is the ordered candidate list the twin produces, $\alpha$ is the target miscoverage and $\delta$ the allowed failure probability. A twin that ignores wake-up transients underestimates $R$ for aggressive, low-energy settings. Those settings get placed early in the order, fail on hardware, and push certification toward expensive settings. Our claim is that modelling three physical effects closes most of the resulting "certified energy gap" on temporal event data: onset truncation from start-up latency, state loss from power gating, and spike corruption during supply settling. The claim also covers co-designing the spiking network to tolerate those effects.
Why These Three Directions Meet Here
The wake-up receiver literature supplies the motivation and the hardware regime. Decoupling channel listening from data reception cuts WLAN communication energy by up to 96% (Low-Power and Low-Delay WLAN Using Wake-Up Receivers). Nano-power wake-up receivers at 600 nW let a whole node sleep at 1 µW (SmarTEG). Spreading codes push wake-up sensitivity to −122 dBm (High sensitivity wake-up radio using spreading codes). None of these systems host a spiking processor, and all of them treat what happens after the interrupt as the host's problem.
The neuromorphic literature explains why power gating is necessary rather than optional. On TrueNorth running DvsGesture, 134.3 mW of the 178.8 mW total is idle leakage, and only 44.5 mW is active gesture computation (A Low Power, Fully Event-Based Gesture Recognition System). Event sparsity alone therefore does not deliver near-zero idle power. The closest sub-microwatt wake-up mechanism in neuromorphic hardware is the 220 nW, sub-0.5 ns common-mode wake-up of a wired address-event link (Qiao & Indiveri). That scheme is galvanic and does not carry over to a radio receiver or to a gated compute core that must restore state.
The simulation-to-silicon literature supplies the method. Air Learning profiled the latency distribution of an embedded processor and injected it into training. This shrank the flight-energy gap between simulation and a real drone from 15.49% to 0.1%. We reuse that idea, but inject wake-up start-up delay rather than inference delay. Rasch et al. showed that injecting measured hardware noise into the forward pass lets analog in-memory networks recover iso-accuracy. They also found that noise at the output interface hurts far more than noise on the weights. We treat the wake-up interface as that kind of output-side corruption of the spike stream. Walker et al. showed that power models grounded in physics ($P \propto V^2 f$ plus temperature-dependent leakage) generalise where counter-fitting models do not, and that argues for a physically parameterised twin over a black-box one. RSVP constrains the design: near-threshold operation causes timing errors that simulators miss. We therefore model the settling window as a period of elevated spike corruption, not as dead time only.
Finally, Is Neuromorphic MNIST Neuromorphic? exposes a weakness in the closest prior work. A frame-based network on time-collapsed N-MNIST reaches 99.23%, so N-MNIST carries almost no temporal information. Chen et al.'s evaluation on N-MNIST alone could not have revealed onset-truncation effects. Spiking Heidelberg Digits (SHD) is different: a linear SVM on spike counts gets about 55%, while a recurrent spiking network reaches 83.2% (The Heidelberg Spiking Data Sets). On data like SHD, timing matters and losing the onset should hurt.
Research Program
The first component is a hardware characterisation rig. A physical wake-up receiver front-end, of the envelope-detector class used in SmarTEG, sits in the always-on domain and drives the power-enable line of a gated processor running a spiking network. A transmitter replays spike streams from benchmark recordings, prefixed by a wake-up packet. We measure four things with an oscilloscope and current probe across supply voltage and temperature: the distribution of delay from wake-up signal to first correctly ingested spike, the inrush current and rail dip, the fraction of spikes corrupted or dropped during settling, and the energy of a complete wake–infer–sleep cycle. We aim for a few thousand wake-ups per condition, which is enough to fit tails. Following Walker et al., the energy model is parameterised physically, with voltage, frequency and temperature terms, rather than as a per-event constant.
The second component is the transient-aware twin. It extends the channel simulation used by Chen et al. with a stochastic wake-up process. Each wake-up samples a start-up delay from the measured distribution and discards spikes that arrive within it. Spikes in the settling window are then flipped or dropped at the measured rate. Membrane potentials are reset to zero whenever the processor was gated, unless a retention option is enabled, and that option is charged its measured retention power. The twin is fitted on one set of conditions and validated on held-out voltage/temperature points, so that its fidelity is itself an empirical result.
The third component is wake-up-aware training. Spiking networks are trained with surrogate gradients on streams passed through the twin's transient process, as in Air Learning's latency injection. The network therefore learns to classify from truncated, reset-initialised streams. We also give the network a learnable input delay line in the spirit of TSkips, which showed that explicit temporal delays improve accuracy on DvsGesture and SHD. The receiver can then emphasise later, more robust parts of the stream.
The fourth component is certified calibration. The DT-LTT procedure of Chen et al. is run unchanged, except that the candidate ordering comes from the transient-aware twin. The guarantee is untouched, so every energy saving we report is at a certified error rate.
Baselines
The naive reference point is an always-on receiver. The spiking processor and main radio never sleep, so there are no transients, but leakage is paid continuously. Amir et al. report 178.8 mW total and 134.3 mW leakage for DvsGesture on TrueNorth at 96.49% ten-class accuracy. On our own platform we will measure the always-on power ourselves, since those figures belong to different silicon. Beating this reference means lower energy per sensing window at equal certified miscoverage, over realistic event rates. We will also report the event rate at which gating stops paying off.
The strongest existing method is DT-LTT as published by Chen et al. It uses a twin that models fading channels but assumes instantaneous, lossless wake-up, and we run it on the same hardware, datasets, network architectures and hardware calibration budget. There is no reported result found on Lacuna for DT-LTT energy or set size on any temporal dataset. The paper reports only that it "consistently met target miscoverage losses" on N-MNIST. We therefore re-run it ourselves. Beating it means strictly lower certified energy per window, or smaller average prediction sets at equal energy, at the same $\alpha$ and $\delta$ across at least five random calibration splits. For accuracy we compare against the published references for unconstrained networks: 83.2% for the best recurrent spiking network on SHD, and 50.9% for the recurrent spiking network and 75.0% for the LSTM on SSC (Cramer et al.). These tell us how much accuracy truncation costs in absolute terms.
Two ablations isolate the key ingredient. The first uses the transient-aware twin for candidate ordering but trains networks on clean streams. This separates the value of better calibration from the value of robust training. The second uses a latency-only twin, which models start-up delay but ignores state reset and settling corruption. This is a direct port of the Air Learning recipe and tests whether the extra physics matters. Both ablations share the same hyperparameter search budget as the full system.
Why not just…?
Why not just buffer spikes or add a guard interval, so nothing is lost during start-up? This is the most serious alternative, and it may win in some regimes. The transmitter can wait a fixed guard time after the wake-up signal, or a small always-on FIFO can hold incoming events. Either way, no onset spikes are lost. The costs are added latency, a worst-case guard length, and always-on buffer power. How large those costs are depends entirely on the tail of the start-up distribution, which no paper in our file reports. We therefore treat the guard interval as a first-class design variable inside the calibration. We will report where guard-only designs match ours. If a guard set from the measured 99th-percentile delay matches our certified energy at acceptable latency, the honest conclusion is that the contribution is the measured twin, which is needed to set the guard, and not the robust training.
Why not just certify directly on hardware and skip the twin? Learn-Then-Test's guarantee holds without a twin. Without a good candidate ordering, however, testing many threshold combinations spends hardware samples and statistical power. We will sweep the hardware calibration budget from small to large. We expect the transient-aware twin's advantage to be largest at small budgets and to vanish as the budget grows. That crossover point is a practical result in itself.
Why not just keep the processor on with state retention and avoid resets? Retention removes state loss but keeps part of the leakage that motivated gating. Amir et al.'s 75% leakage share shows how much is at stake. Retention is one of the twin's modes and is charged its measured power, so the experiments decide the question per event rate rather than by assumption.
Why not just inject latency jitter, as Air Learning did? This is our second ablation. If the latency-only twin closes as much of the certified energy gap as the full model, the reset and settling terms are unnecessary, and we would say so.
Proposed Experiments
We will evaluate on SHD and Spiking Speech Commands (SSC) as auditory temporal benchmarks and on DvsGesture as a visual temporal benchmark. N-MNIST serves as the control, since Iyer et al.'s analysis implies it carries little onset-sensitive information. The primary metric is certified energy per sensing window at $\alpha \in \{0.05, 0.1, 0.2\}$ and $\delta = 0.1$. Secondary metrics are average prediction-set size, end-to-end latency, the realised miscoverage on held-out hardware runs (which must stay below $\alpha$), and the twin's predicted-versus-measured energy error, reported as mean absolute percentage error in Walker et al.'s style.
The decisive check is the interaction between dataset and twin. The certified energy gap between DT-LTT and our method should be large on SHD and DvsGesture and close to zero on N-MNIST. If the gap is equally large on N-MNIST, the gain comes from something other than temporal onset information, and the mechanism we claim is wrong. If the gap is absent on SHD, wake-up transients do not matter at these time scales and the proposal fails. A second check is out-of-distribution fidelity. At held-out supply voltages and temperatures, the transient-aware twin's candidate ordering should still certify cheaper settings than the idealised twin. Otherwise the twin is overfit to the characterisation rig. The third check sweeps the event rate, from sparse industrial triggers like SmarTEG's to near-continuous streams. The curve should show where gating with our calibration beats always-on operation and where it stops helping.
References
- Neuromorphic Split Computing With Wake-Up Radios: Architecture and Design via Digital Twinning
- Low-Power and Low-Delay WLAN Using Wake-Up Receivers
- SmarTEG: An Autonomous Wireless Sensor Node for High Accuracy Accelerometer-Based Monitoring
- High sensitivity wake-up radio using spreading codes: design, evaluation, and applications
- A Low Power, Fully Event-Based Gesture Recognition System
- A Clock-Less Ultra-Low Power Bit-Serial LVDS Link for Address-Event Multi-chip Systems
- Air Learning: a deep reinforcement learning gym for autonomous aerial robot visual navigation
- Hardware-aware training for large-scale and diverse deep learning inference workloads using in-memory computing-based accelerators
- Accurate and stable empirical CPU power modelling for multi- and many-core systems
- RSVP: Soft Error Resilient Power Savings at Near-Threshold Voltage Using Register Vulnerability
- Is Neuromorphic MNIST Neuromorphic? Analyzing the Discriminative Power of Neuromorphic Datasets in the Time Domain
- The Heidelberg Spiking Data Sets for the Systematic Evaluation of Spiking Neural Networks
- TSkips: Efficiency Through Explicit Temporal Delay Connections in Spiking Neural Networks