nnActive: Data-Driven ML Adaptations

• nnActive is a context-dependent term in machine learning that encompasses active-learning methods for 3D biomedical segmentation, adaptive activation networks, and activation-space anomaly detection.
• In 3D segmentation, nnActive extends nnU-Net with partial annotations and patch-based query selection using uncertainty measures like Predictive Entropy and BALD to improve annotation efficiency.
• Alternate usages involve adaptive per-feature polynomial activations and mutual-information-based nearest-neighbor querying, showcasing flexible, data-driven adaptation strategies.

Searching arXiv for papers using the term "nnActive" to ground the article and disambiguate the topic. nnActive is a context-dependent label in the machine-learning literature rather than a single standardized method. In current usage, it most prominently denotes an open-source active-learning framework for 3D biomedical image segmentation built around nnU-Net and partial-annotation patch querying (Lüth et al., 24 Nov 2025). In other strands of the literature, the same name has been attached to neural networks with adaptive per-feature polynomial activations, to activation-space anomaly detection in follow-up usage of A³, and to information-theoretic nearest-neighbor query selection for metric learning and classification (Jang et al., 2018, Sperl et al., 2020, Nadagouda et al., 2022). The shared motif is data-dependent adaptation: query policies, activation functions, or decision rules are conditioned on intermediate structure rather than fixed a priori.

1. Terminological scope

The term has been used for multiple, technically unrelated constructions.

Usage	Representative paper	Core definition
3D biomedical active learning	(Lüth et al., 24 Nov 2025)	Active-learning extension of nnU-Net with partial annotations and 3D patch-based query selection
Adaptive activations	(Jang et al., 2018)	Neural networks in which an auxiliary activation network predicts per-feature polynomial coefficients
Activation-space anomaly detection	(Sperl et al., 2020)	Alarm network operating on hidden activations, optionally supported by an anomaly generator
Active nearest-neighbor querying	(Nadagouda et al., 2022)	Mutual-information selection of nearest-neighbor similarity queries for metric learning and classification

A further nearest-neighbor usage appears in pool-based active learning in metric spaces, where MARMANN is described as an active nearest-neighbor learner that returns a compressed $1$-NN classifier after an active scale-selection stage and compression-set generation (Kontorovich et al., 2016). Accordingly, the term is best interpreted through the surrounding domain and citation context rather than by name alone.

2. nnActive in 3D biomedical segmentation

In its most specific recent sense, nnActive is an active-learning framework for 3D biomedical segmentation designed to overcome four evaluation pitfalls: restriction to too few datasets and annotation budgets, using 2D models on 3D images without partial annotations, Random baseline not being adapted to the task, and measuring annotation cost only in voxels (Lüth et al., 24 Nov 2025). Its stated solution is a large-scale study spanning four biomedical imaging datasets and three label regimes, an extension of nnU-Net with partial annotations for training and 3D patch-based query selection, Foreground Aware Random baselines, and a foreground-efficiency metric.

The framework is implemented as a drop-in extension to nnU-Net $3$D full-resolution with iterative active learning, patch-wise querying, and training on partial annotations. Query selection is patch-based: for each unlabeled image $X^{(i)}$, voxel-wise uncertainty $U$ is computed from an active-learning criterion such as Predictive Entropy or BALD, patch scores are obtained by averaging uncertainty over candidate patches, and greedy top-$k$ non-overlapping patches are retained before a global top-$n$ selection step (Lüth et al., 24 Nov 2025). Training uses a partial-loss over labeled voxels only,

$$L(\theta)=\sum_{i=1..N}\sum_{(x,y,z)\in S^{(i)}} \ell\!\bigl(f_\theta(X^{(i)})_{x,y,z},Y^{(i)}_{x,y,z}\bigr),$$

with a region-sampling trainer that places the known-voxel patch inside a larger context patch.

The benchmark covers four public 3D datasets: AMOS2022, ACDC, KiTS2021, and MSD Hippocampus. Query patch sizes are dataset-specific, reported as $32\times74\times74$ for AMOS, $4\times40\times40$ for ACDC, $64\times64\times64$ for KiTS, and $3$0 for Hippocampus (Lüth et al., 24 Nov 2025). Five-fold cross-validated ensembling per active-learning iteration is used for more stable uncertainty estimates, and the original benchmark retrains fully at each loop rather than fine-tuning.

3. Query modules, baselines, and annotation-efficiency measures

A central design choice in nnActive is that the query module is modular. Reported baselines include naive Random sampling, improved Random $3$1 FG, improved Random $3$2 FG, Predictive Entropy, BALD, PowerBALD, SoftrankBALD, and PowerPE (Lüth et al., 24 Nov 2025). The Foreground Aware Random strategies exist because standard Random sampling can select background-only patches and thereby distort both measured annotation effort and comparisons to uncertainty-driven querying.

The Random $3$3 FG strategy enforces a fixed fraction of foreground-containing patches. For each queried patch, with probability $3$4, the procedure oversamples foreground by either centering a patch on a random voxel from a random foreground class or centering it on a random foreground-border voxel; otherwise, the patch center is drawn uniformly at random over the image (Lüth et al., 24 Nov 2025). This is intended to alleviate both background-only selections and class-size bias.

Foreground Efficiency, abbreviated FG-Eff, is introduced to replace raw voxel count as the sole cost proxy. Let $3$5 denote the fraction of foreground voxels annotated so far, $3$6 the Mean Dice at that foreground budget, and $3$7 the Dice obtained by training on the full dataset. nnActive fits

$3$8

and defines $3$9 (Lüth et al., 24 Nov 2025). Larger $X^{(i)}$0 indicates faster recovery of full-data performance per annotated foreground voxel.

A later extension added Class-stratified Scheduled Power Predictive Entropy, or ClaSP PE, to address class imbalance and redundancy in early selections (Lüth et al., 20 Jan 2026). For a patch $X^{(i)}$1, predictive entropy is

$X^{(i)}$2

and class-stratified scores are defined by

$X^{(i)}$3

A fraction $X^{(i)}$4 of the batch is selected by equal per-class contribution from top-$X^{(i)}$5 patches, while the remainder is selected by top-$X^{(i)}$6 on the original entropy score. Diversity is injected through scheduled Gumbel perturbation,

$X^{(i)}$7

with

$X^{(i)}$8

4. Empirical findings and methodological controversy

The benchmark results reported by nnActive are deliberately cautious. The large-scale study states four headline findings: while all active-learning methods outperform standard Random sampling, none reliably surpasses an improved Foreground Aware Random sampling; benefits of active learning depend on task specific parameters; Predictive Entropy is overall the best performing active-learning method, but likely requires the most annotation effort; and active-learning performance can be improved with more compute intensive design choices (Lüth et al., 24 Nov 2025). This made strong random baselines a central methodological issue in 3D biomedical active learning.

Dataset-specific behavior is emphasized. On AMOS, which contains many small organs, even powerful active-learning methods and Random $X^{(i)}$9 FG struggle to reliably capture small classes, and Random FG baselines often do best. On ACDC, active-learning gains appear in later loops and greedy methods excel. On Hippocampus, greedy methods such as BALD and Predictive Entropy work well. On KiTS, active learning beats FG-aware Random over all regimes, and the best FG-Eff is reported for PowerPE (Lüth et al., 24 Nov 2025). Smaller query sizes and longer training are associated with better AUBC and Final Dice, but at higher compute cost.

A later paper revisits the same benchmark and claims a more decisive result. Using ClaSP PE within the nnActive framework, it reports that, over $U$0 experimental settings on four 3D biomedical datasets, ClaSP PE is the only method that generally outperforms improved random baselines in terms of both segmentation quality with statistically significant gains, whilst remaining annotation efficient (Lüth et al., 20 Jan 2026). The same study further reports evaluation on four previously unseen datasets without manual adaptation, with all experiment parameters fixed according to predefined guidelines, and concludes that the method robustly generalizes to novel tasks without dataset-specific tuning. Taken together, these results identify a live controversy: whether uncertainty-based active learning can consistently exceed strong, task-adapted random baselines in realistic 3D annotation workflows.

5. nnActive as adaptive per-feature activation networks

A distinct use of the name refers to neural networks with activation networks, in which each pixel, node, and layer is assigned a polynomial activation function whose coefficients are produced by an auxiliary network (Jang et al., 2018). The motivation is that standard deep networks apply a single fixed activation such as ReLU uniformly, whereas the proposed formulation makes activation shape depend on neighboring pixels in convolutional layers or on other nodes in dense layers. The method is presented as exploiting interdependency of features and as generalizing fixed activations, fixed polynomial activations, lateral inhibition, and attention modulation.

For a dense layer $U$1, with activations $U$2, the intermediate output is

$U$3

the activation network predicts polynomial coefficients

$U$4

and the activated output is

$U$5

For convolutional layers, the same idea is applied per spatial location:

$U$6

$U$7

$U$8

If $U$9 or $k$0 are zero and the biases are chosen to match a Taylor expansion, fixed activations such as ReLU, sigmoid, or tanh are recovered as special cases.

The reported implementation uses joint backpropagation through both the main network and the activation network, with $k$1 giving good results in the experiments (Jang et al., 2018). On CIFAR-10, baseline LeNet attains $k$2 accuracy, whereas LeNet-AN with $k$3 attains $k$4 with $k$5 more parameters. On MNIST denoising, baseline U-net reports MSE $k$6, while U-net-AN reports $k$7, approximately $k$8 relative error reduction, with $k$9 more parameters. On CIFAR-10 with VGG16, the baseline reaches $n$0, while VGG16-AN reaches $n$1 with $n$2 more parameters. In this usage, nnActive denotes adaptive activation design rather than active-learning-based data acquisition.

6. Activation-space anomaly detection and nearest-neighbor active querying

Another usage centers on hidden activations as a feature space for anomaly detection. In A³, sometimes referred to as nnActive in follow-up work, a pretrained target network $n$3 produces a hidden activation vector $n$4, and an alarm network $n$5 predicts an anomaly score from $n$6 (Sperl et al., 2020). The alarm network is defined by

$n$7

and is trained with binary cross-entropy on normal and anomalous examples. An anomaly network $n$8 can generate pseudo-anomalies, either by direct Gaussian sampling $n$9 or by a VAE-based perturbation of latent codes. The combined objective is

$$L(\theta)=\sum_{i=1..N}\sum_{(x,y,z)\in S^{(i)}} \ell\!\bigl(f_\theta(X^{(i)})_{x,y,z},Y^{(i)}_{x,y,z}\bigr),$$0

with $$L(\theta)=\sum_{i=1..N}\sum_{(x,y,z)\in S^{(i)}} \ell\!\bigl(f_\theta(X^{(i)})_{x,y,z},Y^{(i)}_{x,y,z}\bigr),$$1 in the reported experiments. On image, intrusion-detection, and financial-fraud benchmarks, A³ reports ROC-AUC values in the range $$L(\theta)=\sum_{i=1..N}\sum_{(x,y,z)\in S^{(i)}} \ell\!\bigl(f_\theta(X^{(i)})_{x,y,z},Y^{(i)}_{x,y,z}\bigr),$$2 to $$L(\theta)=\sum_{i=1..N}\sum_{(x,y,z)\in S^{(i)}} \ell\!\bigl(f_\theta(X^{(i)})_{x,y,z},Y^{(i)}_{x,y,z}\bigr),$$3 and average precision in the range $$L(\theta)=\sum_{i=1..N}\sum_{(x,y,z)\in S^{(i)}} \ell\!\bigl(f_\theta(X^{(i)})_{x,y,z},Y^{(i)}_{x,y,z}\bigr),$$4 to $$L(\theta)=\sum_{i=1..N}\sum_{(x,y,z)\in S^{(i)}} \ell\!\bigl(f_\theta(X^{(i)})_{x,y,z},Y^{(i)}_{x,y,z}\bigr),$$5, with only $$L(\theta)=\sum_{i=1..N}\sum_{(x,y,z)\in S^{(i)}} \ell\!\bigl(f_\theta(X^{(i)})_{x,y,z},Y^{(i)}_{x,y,z}\bigr),$$6 to $$L(\theta)=\sum_{i=1..N}\sum_{(x,y,z)\in S^{(i)}} \ell\!\bigl(f_\theta(X^{(i)})_{x,y,z},Y^{(i)}_{x,y,z}\bigr),$$7 real anomaly labels in the semi-supervised setting.

In metric learning and classification, NNActive or Info-NN denotes active nearest-neighbor queries rather than activation-space modeling (Nadagouda et al., 2022). A query $$L(\theta)=\sum_{i=1..N}\sum_{(x,y,z)\in S^{(i)}} \ell\!\bigl(f_\theta(X^{(i)})_{x,y,z},Y^{(i)}_{x,y,z}\bigr),$$8 of length $$L(\theta)=\sum_{i=1..N}\sum_{(x,y,z)\in S^{(i)}} \ell\!\bigl(f_\theta(X^{(i)})_{x,y,z},Y^{(i)}_{x,y,z}\bigr),$$9 consists of a reference point $32\times74\times74$0 and a candidate set $32\times74\times74$1; the oracle returns the item judged most similar to the reference. Query selection maximizes mutual information,

$32\times74\times74$2

under a Plackett-Luce oracle model

$32\times74\times74$3

where $32\times74\times74$4. The paper further shows that active classification can be reformulated as nearest-neighbor query selection by constructing one nearest labeled example per class around an unlabeled point. Reported results include $32\times74\times74$5 MNIST accuracy with $32\times74\times74$6 labels for Info-NN-3, compared with $32\times74\times74$7 for BatchBALD and MaxEntropy, $32\times74\times74$8 for $32\times74\times74$9-Center, and $4\times40\times40$0 for Random.

A more theoretical nearest-neighbor formulation appears in MARMANN, an active learner in metric spaces that returns a $4\times40\times40$1-NN classifier after active model selection and compression-set generation (Kontorovich et al., 2016). The algorithm first chooses a scale $4\times40\times40$2 with $4\times40\times40$3, then builds a compression set $4\times40\times40$4 with $4\times40\times40$5, and finally returns $4\times40\times40$6. Its reported label complexity is, with high probability,

$4\times40\times40$7

and in the regime $4\times40\times40$8 this yields $4\times40\times40$9 labels. This usage reinforces that, in parts of the literature, nnActive is tied to active nearest-neighbor selection rather than to nnU-Net or adaptive activation functions.