Where machine learning meets light

Before asking which model reconstructs or classifies best, I ask a narrower question: which measurements are reliable enough to interpret at all? In optical feedback and single-pixel imaging, alignment mismatch, feedback fluctuation and speckle decorrelation set the ceiling, and a heavier model does not lift it.

So I work the same way on every dataset. Start from an existing measurement and an existing baseline. Name the bottleneck: noise, artefacts, reconstruction, tracking or subject variability. Add one controlled alternative rather than a new pipeline. Then check whether the final useful output improves, not only the intermediate score, because that is what decides whether the method is worth keeping.

The transferable part of MIRABILIS is the method rather than the hardware: feature extraction on imperfect signals, reliability selection before interpretation, reconstruction under model mismatch, and decisions made under uncertainty.

The same questions come up wherever a physical measurement has to become a number someone can trust. I am particularly interested in biomedical sensing: ECG and EMG under wearable conditions, cardiorespiratory and stress monitoring, and ultrasound-derived measurements. There the useful question is usually which windows, frames or sources can be trusted, rather than which classifier to use.

Compressed sensing and biomedical source separation also share a mathematical structure. Both reconstruct information from limited, mixed or noisy measurements.