Measurements are never quite the truth
Any real detector distorts what it measures: limited resolution, acceptance, and noise turn a true distribution into a measured one. Recovering the former is a deconvolution problem.
Bachelor's thesis
Ordinal classification with neural networks in DSEA+
An animated walkthrough of my bachelor’s thesis on combining ordinal neural-network classification with iterative deconvolution.
From distorted measurements to ordinal unfolding
DSEA+ reframes deconvolution as classification
Rather than invert an often unstable matrix directly, DSEA+ iterates: train a classifier, estimate the spectrum from its confidences, reweight the training data, and repeat until convergence.
Any classifier will do — almost
DSEA+ accepts probabilistic classifiers, but its target bins are ordered. Unlike unrelated color categories, bin 5 is closer to bin 6 than to bin 9 — a distance that common softmax classifiers do not encode.
CORN: rank instead of raw category
CORN casts ordinal classification as chained binary questions: does an event exceed rank k? Its probabilities are linked so predictions remain rank-consistent by construction.
Adapting CORN to fit DSEA+
CORN natively returns one predicted rank per event. DSEA+ needs complete per-class confidence distributions and weighted training samples, so both extensions bridge the methods.
Testing the combination
Bayesian hyperparameter search, cross-validation, and bootstrapped uncertainties test whether respecting ordinality improves the unfolded result over a non-ordinal baseline.
Recover distributions, not individual events
The central task is to infer an underlying spectrum from detector-distorted observations while managing limited data and numerical instability.
Iterative, classifier-based unfolding
DSEA+ turns deconvolution into a data-reweighting loop driven by calibrated class-confidence estimates.
Use the structure of ordered targets
CORN encodes rank order directly, providing a principled alternative to treating every target bin as unrelated.
Want to learn more?
Contact me to get access to the full thesis PDF and to discuss the methods and results in more detail.