Germanium detectors searching for neutrinoless double-beta decay must keep a rare signal while rejecting background events that outnumber it by orders of magnitude. Separating the two by the shape of each detector pulse is a central part of that rejection. The signal deposits its energy at one point in the crystal. Gamma-ray backgrounds scatter several times, and surface alpha particles release their charge late. Today the field uses one of two tools. The first is a few hand-designed one-number summaries of each pulse, each cut separately. The second is a learned neural network that compresses the pulse and feeds it to a black-box classifier. The hand-designed cuts are easy to audit and to correct when the detector drifts, but they leave information unused. The neural networks use more of the information, but their outputs are not probabilities. They are also hard to correct when real hardware differs from the data they were trained on.
We propose a middle path built on the oldest trick in probabilistic classification, the naive Bayes assumption. Naive Bayes assumes that, once you know the class, every feature is statistically independent of every other, so the class likelihood factorizes into a product of one-dimensional pieces. Real features almost never satisfy this, which is why naive Bayes usually ranks events well but reports badly exaggerated probabilities. Our idea is to stop hoping the assumption holds and instead train it in. We train a pulse autoencoder (a network that compresses a waveform to a few numbers and reconstructs it) and penalize statistical dependence between its latent coordinates within each class. If that works, the discriminant becomes a sum of one-dimensional log-likelihood ratios. Each term can be checked, recalibrated on the detector's standard calibration peaks, and shifted for systematic uncertainty independently. A dense classifier offers none of these operations.
Core Claim
A latent representation that is approximately independent within each class turns the naive Bayes score from a ranking heuristic into an honest likelihood ratio. This should buy three things that pulse-shape classifiers currently lack. The first is calibrated signal-acceptance estimates. The second is recalibration under detector or simulation mismatch that uses only one-dimensional quantile matching on a small calibration set. The third is a per-coordinate account of where rejection comes from. The bar is that this has to hold without giving up rejection relative to the dense-classifier approach of Deep learning based pulse shape discrimination for germanium detectors. If the independence penalty costs rejection that the transfer and calibration gains do not repay, the idea fails, and the experiments are designed to show that.
Positioning Against Prior Work
The closest work is the autoencoder study of Holl et al. It compresses 256-sample germanium current pulses into a 7-dimensional latent code. A 141-parameter classifier trained on the double-escape peak (a single-site proxy) versus the full-energy peak (a multi-site proxy) then matches the classical amplitude-over-energy parameter $A/E$. That paper supplies our architecture and our benchmark. We differ in what we ask of the latent space. Holl et al. leave the latent coordinates freely correlated and read them with a dense network. We force within-class independence and read them with a factorized generative model.
The hand-crafted line motivates the factorized design. α-event characterization and rejection in point-contact HPGe detectors applies $A/E$ and a delayed-charge-recovery slope as separate one-dimensional cuts. It rejects essentially all surface alphas more than 3 mm from the detector center at 0.2% bulk loss, but performs worse near the point contact. Separate cuts are the deterministic cousin of naive Bayes, and their auditability is exactly what we want to keep. In other detector types, Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment reaches roughly $10^7$ rejection with sequential cuts at only about 20% signal efficiency. Implementation of the Random Forest method for the Imaging Atmospheric Cherenkov Telescope MAGIC shows that cutting correlated variables one at a time discards signal. These results constrain us: naive Bayes on raw, correlated physics features should be expected to fail, which is why independence must be engineered into the representation rather than assumed.
The independence literature tells us when factorization is safe. Re-evaluating the need for Modelling Term-Dependence in Text Classification Problems finds that modeling pairwise dependence rarely beats independence baselines in text, except for very sparse classes. Rare-event physics is exactly such a sparse-class regime, so we treat that exception as a warning and test for it directly. Naive Feature Selection: Sparsity in Naive Bayes shows that under independence the best $k$-feature subset follows from sorting per-feature terms. We borrow that logic, adapted to continuous coordinates by ranking per-coordinate divergence, to choose which latent coordinates to keep from a wider bottleneck. The paper's closed-form proof covers Bernoulli and multinomial models, so for continuous coordinates the sorting rule is a heuristic we validate empirically, not a guarantee.
Research Program
The model is a convolutional autoencoder in the style of Holl et al. with a latent code $z \in \mathbb{R}^d$, where $d$ is between 7 and 32. It is trained on labeled calibration waveforms $x$ with weak labels $y$ (signal-like or background-like) using
$$\mathcal{L} = \mathbb{E}\,\|x - \hat{x}\|^2 + \lambda \sum_{c} \mathrm{Dep}\big(z_1,\dots,z_d \mid y=c\big),$$
where $\hat{x}$ is the reconstruction and $\mathrm{Dep}$ is a within-class dependence estimate. We will compare a kernel independence statistic (the Hilbert–Schmidt independence criterion, summed over coordinate pairs) with a total-correlation estimate. The weight $\lambda$ trades reconstruction against independence.
After training, we fit one-dimensional class-conditional densities $p_j(z_j \mid c)$ by kernel density estimation or histograms. We score each event by
$$s(x) = \sum_{j=1}^{d} \log \frac{p_j(z_j \mid S)}{p_j(z_j \mid B)},$$
where $S$ and $B$ denote signal and background. When the independence penalty succeeds, $s$ approximates the true log-likelihood ratio, so it is optimal for ranking by the Neyman–Pearson argument and calibratable by Bayes' rule with a known prior.
Three things are then studied. The first is the rejection–independence trade-off as $\lambda$ varies, with residual within-class dependence measured on held-out data. The second is calibration: whether the acceptance predicted from the product of one-dimensional marginals matches the measured acceptance at the operating point. The third is transfer. When moving to a new detector or run, we refit only the one-dimensional marginals, or quantile-match each coordinate, using the few thousand calibration-peak events available. We compare this against retraining or fine-tuning dense models on the same events.
The labels are noisy. The double-escape peak contains some multi-site events, and the full-energy peak contains single-site events. Generative marginals can be fitted as two-component mixtures to account for this, which is a further point of contrast with a discriminatively trained network.
Data access is a real dependency. The thorium calibration waveforms of Holl et al., the TUBE alpha scans of Arnquist et al., and the LEGEND electronics-test data of Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines all belong to collaborations. The program requires an agreement with at least one of them; the file gives no public waveform release.
Baselines
The naive reference point is the classical $A/E$ cut, together with delayed-charge recovery for alphas. The strongest reported numbers come from the file. Holl et al. report that both $A/E$ and their autoencoder reach 90% background (full-energy-peak) rejection at 90% signal (double-escape-peak) acceptance. Arnquist et al. report 100% rejection of passivated-surface alphas beyond 3 mm with 0.2% bulk loss. We recompute $A/E$ on our exact event splits so the comparison is not against a number taken from a paper.
The strongest learned competitor is the Holl et al. autoencoder plus dense classifier, reimplemented with the same encoder width, latent dimension, training events, and hyperparameter search budget as our model. Beating it means two things. Rejection at 90% acceptance must be within statistical error of it or better. In addition, our model must show either clearly lower calibration error, or higher rejection after transfer to a new detector or run when both methods get the same calibration events.
A boosted decision tree on the same latent code is the tree-ensemble standard. Its record is set by Boosted Decision Trees, an Alternative to Artificial Neural Networks, which found it retains about 1.5 times less background than a neural network at 50% efficiency on MiniBooNE particle identification. It is tuned with the same budget.
Two ablations isolate the key ingredient. The first applies naive Bayes to the unpenalized latent code ($\lambda = 0$); it tests whether independence must be trained in at all. The second fits an additive logistic model, a discriminative sum of one-dimensional functions, on the penalized code. It tests whether the benefit comes from the generative likelihood-ratio reading or merely from additivity.
Why not just…?
Why not just calibrate the dense classifier with isotonic regression or Platt scaling? That does fix calibration on the data it is fitted to, and we include it as a baseline. It does not fix transfer. When the detector changes, a single monotone remap of the output cannot undo a shift that moves different latent directions differently. The factorized model can correct each coordinate separately. The transfer experiment below settles which approach recovers more rejection per calibration event. If post-hoc calibration plus fine-tuning matches us there, our claim reduces to interpretability alone, and we will say so.
Why not just fit a full conditional density, such as a normalizing flow, on the 7-dimensional latent and take the exact likelihood ratio? This is the most serious threat. In seven dimensions a flow is trainable, and it avoids the independence penalty's rejection cost entirely. Its weakness is the regime that matters here: a few thousand calibration events per class on a new detector, and a need to shift individual directions for systematic uncertainty. Refitting a joint density on that little data, or editing it one direction at a time, is much harder than doing so for one-dimensional marginals. We will run per-class flows on the same latent as a baseline and compare them as a function of calibration-set size. If flows match us at every size, the factorization is not earning its place.
Why not just run naive Bayes or separate cuts on the hand-crafted physics parameters? Those features are correlated and depend on energy, as the MAGIC and PandaX-III results show, and naive Bayes on them double-counts evidence. We include this as a baseline. We expect it to lose rejection and to be badly miscalibrated, which is the problem our training penalty addresses.
Why not just apply independent component analysis to the raw waveforms? Independent component analysis makes coordinates independent across the whole dataset. Naive Bayes needs independence within each class, and data from a two-class mixture is generally not independent within either class when its overall components are independent. We test this as an ablation by replacing the encoder with independent component analysis.
Proposed Experiments
The first experiment reproduces the Holl et al. benchmark on thorium calibration data. We report rejection at 90% acceptance against the penalty weight $\lambda$, alongside residual within-class dependence and expected calibration error. Calibration error is measured on held-out mixtures with known signal fractions, including fractions far below those seen in training. This probes the extreme-imbalance regime in which naive Bayes posteriors are known to collapse toward 0 and 1.
The second experiment, which is the decisive one, tests transfer. We train on one detector or run and deploy on another. Our model may refit only its one-dimensional marginals, using between $10^2$ and $10^4$ calibration events. The dense, tree, and flow baselines may fine-tune or recalibrate using the same events. Success means our model retains rejection at fixed acceptance better than all of them at the smaller calibration sizes, and that its predicted acceptance matches the measured acceptance within statistical error.
The third experiment extends to surface alphas using the TUBE scan data. It asks whether factorized latent scoring improves rejection near the point contact, where the one-dimensional cuts of Arnquist et al. weaken. Bulk loss must stay at or below 0.2%.
The fourth experiment probes the sparse-class warning raised by Banerjee et al. We subsample the minority class down to tens of events and track where the factorized model falls behind the flow. That point marks the empirical boundary between where independence suffices and where dependence must be modeled.
The fifth experiment tests coordinate selection. With a 32-dimensional bottleneck, we choose $k$ coordinates by sorting per-coordinate divergence, following the selection logic of Askari et al. We compare against exhaustive search on small $d$ to check that sorting picks the right coordinates.
References
- Deep learning based pulse shape discrimination for germanium detectors
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment
- Implementation of the Random Forest method for the Imaging Atmospheric Cherenkov Telescope MAGIC
- Re-evaluating the need for Modelling Term-Dependence in Text Classification Problems
- Naive Feature Selection: Sparsity in Naive Bayes
- Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines
- Boosted Decision Trees, an Alternative to Artificial Neural Networks