Many event records, such as earthquakes, infections and sensor alarms, come from monitoring that covers only part of space and time, and adding coverage is expensive. A model of such data is a spatiotemporal point process: a rate function $\lambda(s,t)$ giving how many events to expect per unit area per unit time. Fitting one requires integrating that rate over whatever region was observed. Recent work trains a neural network to represent the antiderivative of the rate, a function whose mixed derivative is the rate. With such a network, the expected count in any space-time box is an exact difference of a few corner values rather than a noisy Monte Carlo sum. We propose to use this exactness for a different purpose. We treat the antiderivative network as the engine of an active-monitoring loop. Each query is a box of space and time to observe. Each box is scored by how much the model disagrees with itself about the expected count there, after removing the unavoidable Poisson noise in that count.
The central claim is that the box, not the point, is the natural unit of both observation and acquisition in this setting. An antiderivative network makes box-level uncertainty cheap and exact at every box size. An ensemble of such networks gives a clean split of uncertainty into a learnable part and a noise part through the law of total variance. Pointwise uncertainty sampling cannot provide this split, and it cannot score thousands of multi-scale candidate boxes per round without heavy quadrature. If this works, the result is the first active sensing method for neural point processes, and it would test whether exact integration matters for deciding what to measure next as well as for fitting.
Background and Gap
Antiderivative ("automatic integration") networks were introduced for volume rendering by AutoInt. That work fits the derivative network to the signal and then reads integrals off the shared-weight antiderivative network. It reports that smooth activations such as Swish are needed for stable training. Auto-NPP carried this approach to temporal point processes, with a monotone cumulative-intensity network and exact likelihoods. AutoSTPP extended it to space-time with ProdNet, a sum of products of one-dimensional networks, so that integrals over boxes stay tractable. AutoSTPP shows that Monte Carlo integration flattens sharp intensity peaks, and it reports a test log-likelihood of 2.62 versus 2.10 for a Monte Carlo baseline on New Jersey COVID-19 data. These papers supply our model class. All of them fit static, fully observed event logs and give no uncertainty over the intensity. The survey Neural Spatiotemporal Point Processes: Trends and Challenges motivates the setting by stressing that event collection is costly and biased by where observers happen to be.
The closest connecting work does something else. Bayesian Stein Networks put a Laplace posterior over a single integral of a function that can be queried at any point, and they leave active acquisition open. Their problem is one global integral. Ours is a field of counts over many overlapping regions, observed as aggregated event sets rather than as point evaluations. Neural Control Variates with Automatic Integration uses antiderivative networks to reduce Monte Carlo variance in PDE solvers, but its samples are drawn passively. Gaussian-process active quadrature (Adaptive Batch Sizes for Active Learning, Domain-Agnostic Batch Bayesian Optimization via Bayesian Quadrature) does select queries to shrink integral variance. It does so with cubic-cost kernels and point queries, and it has no point-process likelihood.
From active learning we take two lessons. MC Dropout active learning for regression shows that cheap model-disagreement scores beat both random and Greedy Max-Min space-filling selection. ALFLTP shows that selecting by maximum likelihood loss or maximum uncertainty chases unlearnable noise, and that a median-aleatoric rule is more stable. That negative result constrains our design: the acquisition score must not reward regions that are merely busy.
Core Claim and Mechanism
Train an ensemble of $M$ AutoSTPP-style antiderivative networks $F_m$ on the events seen in the currently observed set $O$, a union of boxes. The likelihood for partial observation needs $\int_O \lambda$. With antiderivatives this integral is exact: the sum over boxes in $O$ of signed corner evaluations of $F_m$, with $2^3 = 8$ corners per box in space-time. For a candidate box $R$, let $\Lambda_m(R)$ be member $m$'s expected count. The count $N_R$ we would observe is Poisson given the true rate. The law of total variance then gives
$$\mathrm{Var}(N_R) = \underbrace{\mathbb{E}_m[\Lambda_m(R)]}_{\text{Poisson noise}} + \underbrace{\mathrm{Var}_m[\Lambda_m(R)]}_{\text{model disagreement}}.$$
Only the second term can be reduced by observing $R$. We score boxes by the ratio of disagreement to noise, divided by the box's monitoring cost $c(R)$ (proportional to area times duration):
$$a(R) = \frac{\mathrm{Var}_m[\Lambda_m(R)]}{\mathbb{E}_m[\Lambda_m(R)] \, c(R)}.$$
Under a Gaussian approximation, the information gained about the count grows with this ratio. The ratio replaces ALFLTP's median heuristic with a principled correction. A dense, well-understood cluster has a large Poisson term and a small disagreement term, so it is not chased.
Each candidate costs $8M$ network evaluations whatever its size. That lets us score a full multi-scale hierarchy of boxes, so the loop chooses both where to observe and at what resolution. In each round we select a batch greedily, discounting overlapping boxes. This addresses the clustering problem of myopic acquisition that PixelPick handled with a diversity heuristic.
One real complication must be stated up front. In self-exciting processes, unobserved events still trigger observed ones. Phase one therefore conditions only on observed history and treats the missing triggers as model error. We measure the resulting bias on synthetic data with known ground truth before trusting results on real data.
Research Program
We will build three pieces. The first is an AutoSTPP ensemble whose likelihood handles masked observation sets. The second is the box-hierarchy scorer. The third is a simulated monitoring protocol. In that protocol, a dataset's full space-time domain is hidden except for a small seed fraction, and each round "deploys a sensor" by revealing all events in the chosen boxes until a fixed fraction of the domain-cost budget is used.
The analysis will check two things. First, it will check whether ensemble disagreement over $\Lambda(R)$ is calibrated: on held-out boxes, does the predicted count variance match observed squared errors? Second, it will check how calibration degrades with Swish versus other activations, given AutoInt's stability finding. A temporal-only version on Auto-NPP's synthetic processes comes first, because the true intensity is known there.
Baselines
The naive reference is random box selection under the same cost budget, with the same ensemble, training schedule and hyperparameter search. A second reference is space-filling selection: a regular grid sweep and Greedy Max-Min over box centers. Tsymbalov et al. found this kind of selection weaker than uncertainty sampling for regression, but it is the obvious practitioner choice for sensor placement.
The strongest competitor is pointwise uncertainty sampling in the style of MCDUE (Dropout-based Active Learning for Regression). It uses the same ensemble, scores candidate boxes by Monte Carlo averages of the pointwise intensity variance, and is given the same wall-clock acquisition budget as our scorer. That budget is the fair constraint, because it decides how many Monte Carlo samples each box receives. This baseline isolates the question of whether exact box-level variance, including correlations within the box, beats summed pointwise variance.
The key ablation keeps everything but drops the noise correction. It scores boxes by total count variance, or by maximum held-out negative log-likelihood, which mirrors the rules ALFLTP found unstable. A second ablation replaces the antiderivative model with a Monte Carlo-integrated neural intensity model for both fitting and scoring. This tests whether AutoSTPP's fitting advantage carries over to the active setting.
The final numbers to beat come from AutoSTPP. With full data on COVID-19 NJ, AutoSTPP reports test log-likelihood 2.62 against 2.10 for the Monte Carlo baseline. For Japanese earthquakes the file gives no reported number on Lacuna for AutoSTPP or Auto-NPP log-likelihood. No reported result was found on Lacuna for any active-learning method on point processes. Beating a baseline means the method reaches its final held-out log-likelihood with a significantly smaller revealed budget, or reaches a higher log-likelihood at equal budget, across at least five seeds. The full-data AutoSTPP result serves as the ceiling.
Why not just…?
Why not just use pointwise MC Dropout or ensemble variance on the intensity and sum it over the box? This is the real threat. If intensities are smooth and boxes are few, it may match our method. Two things argue against it. AutoSTPP's evidence shows that Monte Carlo sums blur sharp peaks, and those peaks are exactly where disagreement concentrates. Summed pointwise variance also ignores the correlation of errors within a box. The equal-wall-clock comparison above settles the question. We expect our advantage to grow with the number and range of scales of candidate boxes, and we will report where it vanishes.
Why not just use Gaussian-process Bayesian quadrature or a Gaussian-process count model? GP active quadrature handles point queries of a function with cubic scaling. The BSN paper reports GP quadrature failing around $10^4$ points, while neural integrators reached about $10^6$ and did better on discontinuous Genz functions. A GP also lacks the self-exciting structure of AutoSTPP. We will include a small-scale GP count surrogate on the temporal synthetic tasks, where it is feasible, to show where it stops being competitive.
Why not just sweep space evenly? Space-filling designs are robust and need no model. The MCDUE comparison against Greedy Max-Min suggests they waste budget on quiet regions. The grid and Max-Min baselines test this directly, and a tie would mean the uncertainty machinery is not worth its cost.
Why not just pick the most uncertain or worst-fit boxes? ALFLTP found maximum-NLL selection unstable because it tracks noise. Our total-variance ablation tests the same failure in point processes. We add a synthetic stress test that plants a high-rate homogeneous Poisson region, which is easy to learn but full of noise. The noise-corrected score should leave that region alone, and the uncorrected one should not.
Proposed Experiments
The temporal stage uses Auto-NPP's Shaky Hawkes, Delayed Peak and Shift Hawkes processes, with time windows as queries. It measures intensity recovery error (MAPE against the known truth) as a function of revealed budget, plus the calibration of predicted count variance. The spatiotemporal stage uses Japanese earthquake records and COVID-19 NJ, with seed fractions of 5 percent and budgets up to 40 percent of domain cost. It reports full-domain held-out log-likelihood curves, the budget needed to reach 95 percent of the full-data log-likelihood, and acquisition wall-clock time.
The decisive checks are three. Exact box scoring must beat Monte Carlo-summed pointwise scoring at equal compute. The noise-corrected score must avoid the planted noise region while total variance does not. The bias from conditioning only on observed history, measured on synthetic self-exciting data, must stay small relative to the gains. If the first check fails, the antiderivative network offers no acquisition advantage and the contribution reduces to the noise-corrected score alone.
References
- AutoInt: Automatic Integration for Fast Neural Volume Rendering
- Automatic Integration for Fast and Interpretable Neural Point Processes
- Automatic Integration for Spatiotemporal Neural Point Processes
- Neural Spatiotemporal Point Processes: Trends and Challenges
- Baysian numerical integration with neural networks
- Neural Control Variates with Automatic Integration
- Adaptive Batch Sizes for Active Learning: A Probabilistic Numerics Approach
- Domain-Agnostic Batch Bayesian Optimization with Diverse Constraints via Bayesian Quadrature
- Dropout-Based Active Learning for Regression
- Privacy-Preserving and Uncertainty-Aware Federated Trajectory Prediction for Connected Autonomous Vehicles
- All you need are a few pixels: semantic segmentation with PixelPick