How Can Smart Sampling Make AI More Trustworthy?

Compressive learning (CL) is a new AI frontier: it lets models reason directly on compressed data, saving time and energy. Until recently, CL solutions had to reconstruct the original image to compensate for lost information, a costly step that didn’t always remove uncertainty.

The MCS‑TCL project takes a different route. Using Multi‑functional Compressive Sensing Sampling, sampling, compression, and feature extraction happen in a single operation. In practice, the sensor captures a signal already "ready" for the model, preserving the most useful information for class discrimination. This makes the system faster and less dependent on high sampling rates: with only 6 % of the original data, MCS‑TCL outperforms traditional methods.

To make predictions more trustworthy, the authors added evidential deep learning (EDL). EDL not only gives an answer but also estimates how confident the model is, separating well‑supported predictions from doubtful ones. The result is a much smaller model (85.76 % fewer parameters) that, besides being accurate, provides uncertainty estimates—ideal for resource‑constrained devices like drones or IoT sensors.

In short, MCS‑TCL shows that efficiency and reliability can go hand in hand, paving the way for lighter AI that knows its own limits.

The core of MCS‑TCL lies in the synergy between compressive sensing (CS) and convolutional neural networks (CNN). In CS, a signal σ → ρ is acquired via a measurement matrix Φ, typically random, reducing dimensionality while preserving the Restricted Isometry Property (RIP). MCS‑TCL extends this by introducing a matrix Χ designed to be orthogonal to the CNN filters, ensuring that sampling and feature extraction operations are mathematically compatible.

During training, the model simultaneously optimizes Χ and the CNN weights through back‑propagation, minimizing a composite loss: L = L_{CE} + λ L_{EDL}. L_{CE} is the standard cross‑entropy, while L_{EDL} stems from evidential deep learning, based on a Dirichlet distribution α that models evidence for each class. The Dirichlet variance provides a measure of epistemic uncertainty, useful when compressed data are insufficient or ambiguous.

A crucial aspect is MCS‑TCL's ability to operate at extremely low sampling rates (6%). Thanks to the co‑design of Χ and the CNN, discriminative information is concentrated in the most energetic coefficients, reducing critical information loss. Moreover, the 85.76% parameter reduction is achieved via structured pruning and 8‑bit quantization, maintaining numerical stability thanks to adaptive batch‑normalization.

Experiments on datasets such as CIFAR‑10 and ImageNet showed that MCS‑TCL outperforms traditional CL benchmarks both in accuracy (+3.2% at 6% sampling) and uncertainty calibration (ECE reduced by 45%). These results suggest the approach is promising for edge applications where latency, energy consumption, and decision trust are critical.