When a gamma ray hits the atmosphere, four Cherenkov telescopes photograph the same air shower. When an electron pair crosses a xenon gas detector, five "topological" features are all measured from the same ionization track. Physics analyses often treat these as separate views, one per telescope or one per feature. The tools that machine learning offers for uncertainty-aware fusion of views assume the views are independent witnesses. These tools are evidential fusion methods: each view outputs both a class vote and an explicit "I don't know" mass, and a combination rule pools them. Under the independence assumption, four telescopes looking at one shower count as four pieces of evidence. The fused prediction becomes overconfident in exactly the way that hurts a rare-event search. It claims a background event is rejected with more certainty than the data support, so it underestimates the background leakage at the extreme operating points (one false event in 10⁴ to 10⁷) where discovery claims are made.
We propose redundancy-discounted evidential fusion. Each detector view still produces its own evidence and its own vacuity, which is the share of belief left unassigned when a view has seen nothing useful. This lets a moonlit, saturated, or dead telescope drop out of a given event on its own. The new part is a small learned module that estimates, per event, how much each view's evidence overlaps with the others and shrinks it before pooling. The defining property is simple: feeding the same view in twice must leave the fused opinion unchanged. Vanilla Dempster fusion does not satisfy this, and neither does any method in this literature that we know of. We test whether the property makes the fused uncertainty an honest predictor of background leakage, while keeping the robustness to degraded views that evidential fusion was built for.
Core Claim
Trusted Multi-View Classification (TMC) gives each view a Dirichlet distribution parameterized by non-negative evidence $e^m_k$ for class $k$ from view $m$. It fuses views with Dempster's rule, which for well-separated opinions behaves like adding evidence across views. TMC itself flags the weakness we target: Dempster's rule assumes conditionally independent sources, and correlated views (such as spatial and frequency representations of one image) are double-counted. Later work changes the operator but not this assumption. Liu et al.'s additive aggregation guarantees that fused vacuity never increases as views are added, which is exactly wrong for a duplicated view. TEF searches over fusion operators to build an extra "pseudo view," but it does not discount mutual information. FAML fixes class-prior bias, not view redundancy.
Our replacement is the fused evidence
$$e^{\text{fused}}_k = \sum_{m=1}^{M} w_m\, e^m_k, \qquad w_m = \frac{1}{1 + \sum_{n \neq m} r_{mn}},$$
where $M$ is the number of views and $r_{mn} \in [0,1]$ is a per-event redundancy score between views $m$ and $n$. The score is predicted from their embeddings and, for telescopes, from shower geometry such as the reconstructed impact distance. If two views are exact copies, $r=1$, each gets weight one half, and the fused evidence equals that of a single view. If they are independent, $r=0$ and we recover ordinary cumulative fusion. A degraded view still contributes nothing, because its own $e^m$ goes to zero. The claim is that this discount has three effects relative to undiscounted evidential fusion at matched backbone and compute. It makes the fused Dirichlet calibrated in the far tail of the background distribution. It preserves the per-view drop-out robustness that motivates evidential fusion in the first place (Ashhad et al. document how equal-weight fusion breaks under view corruption). It costs little or no rejection power.
Research Program
The work has three parts. First, we build the fusion layer and its training objective. Each view gets a backbone with an evidential head trained with TMC's digamma-form loss and KL penalty. The redundancy module is a two-layer network on pairs of view embeddings plus physical side information. It is trained jointly with two extra terms. The first is a duplication-consistency term: a view is randomly duplicated or replaced with a noisy copy during training, and the loss penalizes any change in the fused opinion relative to the unduplicated input. The second is a tail-calibration term: on a held-out background-only split, it matches the fused model's predicted probability of misassigning background as signal to the observed leakage, in score bins above the working threshold. The tail term is where the rare-event regime enters. Leakage at fixed signal efficiency is invariant to the training class ratio, but the reported posterior is not. So we also test FAML's adaptive prior as an optional component, rather than assuming imbalance is solved.
Second, we analyze where overcounting actually arises. We measure learned $r_{mn}$ against physically interpretable quantities: angular separation of telescopes from the shower core, and, for xenon tracks, the analytic dependence of track length on blob charge.
Third, we stress-test the method on degradations it was not trained on. These include injected night-sky noise on one telescope, dead camera sectors, a missing telescope, and a shifted zenith angle. They give a partial view of the simulation-to-reality gap that the IACT CNN-GNN study leaves untested.
Baselines
The naive reference points are the classical ones a physicist would use without deep fusion. On the telescope data, this is the Hillas-parameter GNN that the CNN-GNN paper compares against. On the xenon data, it is the sequential cut analysis of PandaX-III topological discrimination, which reports roughly seven orders of magnitude background suppression with signal efficiency above 20%. On HiggsML, it is a single fully connected network on all 30 features. Our method only matters if it does not lose to these at matched signal efficiency.
The strongest existing competitor is the Strong CNN-GNN of Convolution and Graph-based Deep Learning Approaches for Gamma/Hadron Separation. It fuses telescopes implicitly through graph message passing, so it learns cross-view correlation and should not overcount, and it is the published state of the art on this data. That paper reports only that it has higher AUC than the Hillas GNN and an additional 9% false-positive-rate reduction over the Fast CNN-GNN at threshold 0.5. No reported absolute AUC was found on Lacuna for this dataset, so we reproduce it ourselves. To make the comparison fair, we give it the same per-telescope Inception backbone, the same training events, the same hyperparameter search budget, and, critically, the same random telescope-dropout and noise augmentation we use for our duplication term. On HiggsML, the strongest number in the file is DEFE's reported AUC of 0.916 (Deep extreme feature extraction). We quote it as a reference but do not claim to beat it, since our goal there is a controlled correlation test, not a leaderboard.
We also run three uncertainty baselines on the same backbones. These are a five-member deep ensemble of the CNN-GNN, vanilla TMC with Dempster's rule, and Liu et al.'s additive fusion. Two ablations isolate the key ingredient: setting all $w_m=1$ (no discount), and using a single global learned weight instead of per-event redundancy.
Beating the CNN-GNN requires one of two outcomes. The first is matching its background rejection at 50% and 80% gamma efficiency within statistical error on clean data, while showing smaller predicted-versus-observed leakage error in the tail. The second is smaller loss of rejection under unseen telescope degradation. Beating vanilla TMC requires a lower duplication sensitivity and a lower tail-calibration error at equal AUC.
Why not just…?
Why not just train the graph network with random telescope dropout and noise? This is the most serious threat, and it may well match us on raw rejection under degradation. That is why the augmented CNN-GNN is our main baseline, not an afterthought. What it does not give is a per-event, per-telescope account of which views were discounted, or a fused uncertainty that is meant to equal tail leakage. If the augmented CNN-GNN with post-hoc temperature scaling matches our tail-calibration error as well as our rejection, then the redundancy discount adds nothing for physics, and we will report that. The tail-calibration and unseen-degradation experiments decide this directly.
Why not just use a deep ensemble or Monte Carlo dropout? Ensembles express disagreement between models. They do not express the absence of evidence from one view, and they cost several forward passes per event, which matters for trigger-rate processing (sPHENIX real-time AI). TMC reports beating these baselines in accuracy on multi-view benchmarks (for example, by 7.6% on PIE), but not on physics data. So we run the ensemble on our benchmarks at five times the inference cost and compare tail calibration directly, rather than relying on that transfer.
Why not just combine per-view calibrated classifiers by multiplying their likelihood ratios? This is the classical physics shortcut, and Chen and Aickelin's Dempster-Shafer anomaly detector is essentially its belief-function version. It makes the same independence assumption we are trying to remove, so it should overcount at least as much as vanilla TMC. We include it in the duplication test, where we expect it to fail in the same way. If instead it holds up as well as our method there, the evidential machinery is unnecessary.
Proposed Experiments
The first decisive check is the duplication test. We feed each telescope image, or each xenon feature group, twice. Our method's fused vacuity and leakage should stay nearly unchanged. Vanilla TMC, additive fusion, and the likelihood product should all report sharply lower uncertainty.
The second check uses HiggsML 2014 with its 17 low-level and 13 high-level features as two views. The high-level features are computed from the low-level ones, so an honest fusion should be about as uncertain as the better single view. We measure how far each method's fused vacuity falls below that.
The third check is the main IACT study on the H.E.S.S.-style simulation used by the CNN-GNN paper. Data access permitting, we use its 150,000 balanced training events; otherwise we regenerate an equivalent sample with the collaboration simulation chain. We then evaluate on background-heavy test sets at realistic 10⁴:1 proportions. The metrics are background rejection at fixed gamma efficiency, AUC, and the ratio of predicted to observed leakage in the top score bins with bootstrap intervals. We repeat these under single-telescope noise, dead pixels, telescope removal, and a zenith shift.
The fourth check uses PandaX-III topological features from REST simulation ($^{214}$Bi and $^{208}$Tl backgrounds against $0\nu\beta\beta$ signal). This tests strongly correlated within-track views, measured against the published cut analysis at its >20% signal efficiency.
The project succeeds if the discount closes most of the duplication and tail-calibration gap of vanilla evidential fusion while staying within error of the augmented CNN-GNN on clean rejection. It fails, informatively, if the augmented graph network is equally calibrated, which would mean implicit fusion already solves the problem that explicit evidential fusion creates.
References
- Trusted Multi-View Classification
- Trusted Multi-View Deep Learning with Opinion Aggregation
- Trusted Multi-View Classification via Evolutionary Multi-View Fusion
- Fairness-Aware Multi-view Evidential Learning with Adaptive Prior
- Uncertainty-Aware Multi-view Arrhythmia Classification from ECG
- Convolution and Graph-based Deep Learning Approaches for Gamma/Hadron Separation in Imaging Atmospheric Cherenkov Telescopes
- Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment
- Deep extreme feature extraction: New MVA method for searching particles in high energy physics
- Intelligent experiments through real-time AI: Fast Data Processing and Autonomous Detector Control for sPHENIX and future EIC detectors
- Dempster-Shafer for Anomaly Detection