Source-linked AI summary
Wireless Evidence Acquisition for Multimodal Inference Constrained by Temporal Windows
Alessandro Compagnoni, Anup Mishra, Carla Fabiana Chiasserini, Elad Michael Schiller, Petar Popovski
TL;DR
Wireless allocation shapes which time-sensitive sensor observations become evidence for a multimodal prediction, so delivered bits alone do not capture downstream utility. The paper formulates scheduling within a prescribed temporal window as sequential evidence acquisition and develops MIRA to combine conditional evidence gain with channel information and remaining data volume. Across three task types, MIRA outperforms relevance-only and best-channel scheduling, with up to 95% higher classification accuracy against Best-CQI.
Problem
Wireless decisions shape which time-sensitive sensor observations become prediction evidence, so delivered bits alone do not capture downstream utility.
Source: §I · p. 1Method
For LMMSE fusion, MIRA greedily combines conditional evidence gain from residual-error-volume reduction with channel information and remaining data volume, updating relevance as evidence arrives.
Source: §III · p. 3, §V · p. 5Results
Across synthetic classification, human activity recognition, and vehicle-trajectory regression, MIRA outperforms relevance-only and best-channel scheduling, with up to 95% higher classification accuracy against Best-CQI.
Source: §V · p. 5Takeaways & Limitations
MIRA’s gains are more pronounced when modality relevance and channel quality are misaligned, and maximising data need not maximise inference.
Source: §V · p. 5
Abstract
from arXiv · showhide
Multimodal inference in safety-critical applications, such as autonomous driving and robot navigation, requires heterogeneous sensor observations to reach an edge server within a task-prescribed temporal window. Referenced to the event or state being inferred, this window ends at the latest time an observation remains useful for the current decision. Since sensor readiness times, data volumes, wireless channel conditions, and task relevance vary across sensors, maximising network throughput does not necessarily minimise the fused prediction error when the window closes. We cast multimodal uplink scheduling as sequential wireless evidence acquisition and, for a linear minimum mean-square error (LMMSE) fusion model, derive a conditional evidence gain metric from the reduction in residual-error volume. Defined through second-order statistics, this metric applies beyond Gaussian models and coincides with conditional mutual information when the target and prediction errors are jointly Gaussian. We then develop MIRA, a greedy task-driven maximum information-rate allocation policy that combines conditional evidence gain with each sensor's channel state information and remaining data volume, while updating sensor relevance as evidence is acquired. Experiments on synthetic classification, human activity recognition, and vehicle-trajectory regression show that MIRA outperforms both relevance-only and channel-only scheduling. Relative to the former, MIRA improves classification accuracy by up to 70% and reduces regression mean-square error (MSE) by up to 5.5%. Relative to the latter, it requires 58% less acquisition time to attain an 80% target accuracy, while achieving up to 95% higher accuracy and 9% lower regression MSE. These gains are achieved without maximising received-data volume.
I. INTRODUCTION
Wireless decisions shape which time-sensitive sensor observations become evidence for a multimodal prediction, so network throughput alone does not capture downstream utility. The paper frames delivery within a prescribed temporal window as sequential evidence acquisition and introduces MIRA to balance evidence value with wireless acquisition conditions.
I. INTRODUCTION
- In multimodal inference, wireless decisions shape the evidence available for prediction, and evidence usefulness depends on time.Task-oriented communication evaluates wireless design through downstream utility rather than delivered bits alone,.Source: §I · p. 1
- The temporal window of integration and event-inference reliability provide temporal and inferential foundations for evaluating multisensory evidence, [8].The TWI distinguishes task-dependent usefulness from cross-modal simultaneity and causal consistency.Source: §I · p. 1
- The paper addresses the largely unexplored coupling between inferential value and sequential wireless acquisition by balancing each observation’s conditional prediction value against its radio cost.This complements work combining usefulness-based window selection with slot-level wireless control [11].Source: §I · p. 1
- Across synthetic classification, human activity recognition, and vehicle-trajectory regression, MIRA outperforms sum-rate-maximising scheduling, with gains exceeding 95%.Acquiring more data does not necessarily improve inference performance.Source: §I · p. 2
II. SYSTEM MODEL AND PROBLEM FORMULATION
The system acquires heterogeneous sensor observations within an event-referenced usefulness window and seeks a causal wireless allocation policy that minimises fused prediction error when the window closes. Sensor readiness, data volume, and channel conditions shape which evidence can be acquired in time.
II. SYSTEM MODEL AND PROBLEM FORMULATION
- Sensors produce modality-specific predictions from observations completed after sensor-specific integration times, and only data delivered by T contribute to the current prediction.Observations must satisfy the TWI’s event-association, simultaneity, and causal-consistency conditions.Source: §II · p. 2
- The inference node combines complete-observation predictions with an LMMSE rule using offline-estimated means and cross-covariances, including prediction-error dependence.The model covers regression and classification, with one-hot targets and softmax-output predictions for classification.Source: §II · p. 2
- Slot-based scheduling allocates resource blocks among ready sensors with remaining data, whose transmission rates depend on spectral efficiency and allocated bandwidth.Readiness follows sensor integration time; spectral efficiency is selected through link adaptation from channel SNR.Source: §II · pp. 2–3
- For partial reception, missing observation elements are replaced with training-set means for final fusion, while scheduling updates evidence gains using only fully received observations.Data arriving after T are excluded from the current prediction.Source: §II · pp. 2–3
- The allocation policy minimises expected fused prediction error using evidence acquired by deadline T, the end of the task-prescribed window W=[0,T].If inference takes fixed duration T_A, the final prediction is produced by T_max=T+T_A.Source: §II · pp. 2–3
III. THE MIRA POLICY
MIRA values a sensor by the conditional reduction it offers in residual-error volume, then combines that value with the radio service needed to acquire its data. Its slot-based allocation updates the remaining sensors’ evidence gains as observations are completed.
III. THE MIRA POLICY
- The conditional evidence gain is the logarithmic reduction in residual-error volume from completing a sensor observation.The residual-error covariance may be regularized by adding ϵI_C, with ϵ = 10^-8 when the covariance is singular.Source: §III · p. 3
- The gain accounts for redundancy and complementarity, and equals conditional mutual information when the target and prediction errors are jointly Gaussian and ϵ = 0.A sensor’s value changes with the evidence already acquired.Source: §III · p. 3
- MIRA’s priority score favors high conditional evidence gain relative to the radio service needed to complete the remaining data.It scales the value of a complete observation by the fraction of remaining data served by one resource block, assuming task value accumulates proportionally with that fraction.Source: §III · p. 3
- Within each slot, MIRA allocates resource blocks to a selected sensor, completes it if its buffer fits, and updates remaining sensors’ conditional gains.The procedure repeats across slots, adapting to changing wireless state and acquired evidence; prediction uses complete or imputed observations available when the window closes.Source: §III · p. 3
- MIA prioritizes conditional evidence gain without channel state information, whereas Best-CQI prioritizes spectral efficiency without inferential value.Both benchmarks use the same within-slot allocation procedure as MIRA.Source: §III · p. 3
- The stated worst-case complexity per TTI is O(S min{M, S}(SC)^3) when the acquired set expands at every allocation step.Each gain update requires a pseudoinverse of dimension at most SC.Source: §III · p. 4
IV. PERFORMANCE EVALUATION
The evaluation compares MIRA with MIA and Best-CQI across synthetic classification, human activity classification, and vehicle-trajectory regression. Across these settings, MIRA performs best, including when informative sensors have poorer channels.
Evaluation Setup
- The experiments compare all three policies using the same sensor-readiness and channel realisations, with results reported against the acquisition deadline T.They cover synthetic classification, human activity classification, and vehicle-trajectory regression; rank-based SNR deliberately assigns poorer average channels to more informative sensors.Source: §IV · p. 4
Synthetic Multimodal Classification
- 95.11% higher accuracy than Best-CQI and 15.03% higher than MIA are MIRA’s maximum gains in uniform-SNR synthetic classification.For an 80% accuracy target, MIRA and MIA need approximately 500 ms, compared with approximately 1190 ms for Best-CQI.Source: §IV · p. 4, Figure 2 · p. 4
- Under rank-based SNR, MIRA’s accuracy gains reach 35.40% over Best-CQI and 70.10% over MIA in synthetic classification.MIA favors informative but costly sensors, while Best-CQI favors efficiently transmitted but weakly relevant observations.Source: §IV · p. 4, Figure 2 · p. 4
Real-World Human Activity Recognition
- For human activity classification under rank-based average SNR, MIRA’s relative accuracy improvements reach 66.70% over Best-CQI and 17.15% over MIA; MIA matches MIRA at short and moderate horizons.Source: §IV · p. 5, Figure 3 · p. 5
Vehicle-Trajectory Regression
- In vehicle-trajectory regression, MIRA has the lowest MSE throughout, reducing it by up to 9.09% versus Best-CQI and 5.52% versus MIA.The gap between MIA and MIRA narrows at longer horizons.Source: §IV · p. 5, Figure 4 · p. 5
V. CONCLUSIONS
The paper frames multimodal uplink scheduling within a temporal usefulness window as sequential evidence acquisition. It derives an LMMSE-based evidence-gain metric and develops MIRA, which outperforms relevance-only and best-channel scheduling across the evaluated tasks.
Conclusions
- The paper formulates multimodal uplink scheduling within a prescribed temporal window as wireless evidence acquisition, where radio-resource decisions construct the evidence available for prediction.Source: §V · p. 5
- For LMMSE fusion, the paper derives an information-based metric from residual-error-volume reduction and develops MIRA to combine that gain with channel state and remaining data volume.The metric coincides with conditional mutual information under joint Gaussianity.Source: §V · p. 5
- Across synthetic classification, human activity recognition, and vehicle-trajectory regression, MIRA outperforms relevance-only and best-channel scheduling.Against Best-CQI, it achieves up to 95% higher classification accuracy and 9% lower regression MSE, while requiring up to 58% less time to reach 80% target accuracy.Source: §V · p. 5
- Future work will examine nonlinear, learned, or dynamically adapted fusion and lower-complexity evidence-gain approximations for real-time operation on constrained edge nodes.Source: §V · p. 5