AdaptiveAD: Selective Test-Time Adaptation

• AdaptiveAD is an unsupervised anomaly detection paradigm that selectively adapts non-anomalous regions using a lightweight neural implicit representation.
• The approach freezes the base model and optimizes an added MLP adapter with a hybrid loss, ensuring adaptation without erasing true pathology.
• Empirical results demonstrate significant improvements in TPR and reduced reconstruction errors, validating its effectiveness against domain shifts.

AdaptiveAD refers to a family of adaptive, often learning-based, methods across diverse computational domains. The term is most notably used in selective test-time adaptation for unsupervised anomaly detection in medical imaging, end-to-end autonomous vehicle planning, dynamic defense on Active Directory graphs, and advanced distributed optimization. This entry focuses on the key AdaptiveAD paradigm in unsupervised anomaly detection, followed by contextualization in related computational settings.

1. Selective Test-Time Adaptation in Unsupervised Anomaly Detection

AdaptiveAD in unsupervised anomaly detection addresses performance degradation under domain shift, specifically in medical imaging, where reconstruction-based models trained solely on healthy data encounter substantial distribution shifts at inference. Standard approaches such as full fine-tuning of the anomaly detection network at test-time risk catastrophic failure—adapting not only to distribution shift (scanner/protocol/patient population) but also “learning away” actual pathology, thus erasing true anomalies from predictive signals (Ambekar et al., 4 Oct 2024).

AdaptiveAD introduces selective, zero-shot, per-sample adaptation restricted to non-anomalous image regions based on precomputed source-domain feature statistics. Rather than adapting the entire source-trained network $f_{s}$, AdaptiveAD wraps $f_{s}$ with a lightweight neural implicit representation $g_\theta$—architecturally, a two-layer MLP mapping deep feature vectors to pixel/patch intensities.

2. Core Methodology and Neural Implicit Representation

In the AdaptiveAD paradigm, the core model is defined as follows. During inference, $f_{s}$ is frozen. For a test image $x_t$, a pre-trained feature $f \in \mathbb{R}^d$ (from $f_s$’s penultimate layer or independent backbone) is passed to the multi-layer perceptron

$$\hat y = g_\theta(f) = W_2\sigma(W_1 f + b_1) + b_2,$$

with $W_1 \in \mathbb{R}^{h \times d}$, $W_2 \in \mathbb{R}^{p \times h}$, $b_1 \in \mathbb{R}^h$, $b_2 \in \mathbb{R}^p$, and nonlinear activation $\sigma$ (e.g., ReLU). Here, $p$ is the number of pixels in the patch or output vector.

The adaptation step updates only the parameters $\theta$ of $g_\theta$, never those of $f_s$. The test image is decomposed into overlapping patches, features are extracted, and $g_\theta$ reconstructs each patch, which are then recombined to form the adapted prediction $\hat y_t$. The typical $g_\theta$ has ~600K parameters, with a batch-norm-only variant reducing this to 4K for efficiency.

3. Selective Adaptation and Loss Function

AdaptiveAD minimizes a hybrid adaptation loss per test-sample:

$$\mathcal{L}(\theta) = \lambda_r \| x_t - \hat y_t \|^2 + \lambda_s S(f),$$

where the first term enforces pixel-wise or feature-space reconstruction fidelity, and $S(f)$ is a feature-selection penalty restricting adaptation to likely non-anomalous features. $S(f)$ is computed using source-domain per-feature mean $\mu_j$ and variance $\sigma_j^2$. Features $f_j$ that deviate from the healthy envelope (e.g., $|f_j - \mu_j| > k \sigma_j$) are masked, setting $m_j = 1$ in $S(f)$ only if outside the envelope, otherwise $0$.

The penalty

$$S(f) = \sum_j m_j \| W_2 \sigma(W_1 f + b_1)_j \|^2$$

“freezes” portions of the implicit map in regions signaling pathological deviation, preventing adaptation from “fitting out” pathological signal.

4. Training and Inference Protocol

AdaptiveAD operates in a fully zero-shot, sample-wise regime:

1. Initialize $g_\theta$ to its source (unadapted) state.
2. Extract features from $x_t$.
3. Compute the selection mask $m$ based on deviations from feature-wise source statistics.
4. Optimize $\theta$ with gradient descent on $\mathcal{L}(\theta)$ (learning rate $10^{-4}$, typically 50–100 steps).
5. Assemble the adapted output $\hat y_t$ patch-wise.
6. Compute anomaly map $s = |x_t - \hat y_t|$.

The approach requires no target-domain retraining, labels, or sample accumulation. Adaptation time is ≈4 minutes per 2D slice (typical setting), reduced to <1 min with batch-norm tuning.

5. Empirical Results and Evaluation

AdaptiveAD was evaluated on unsupervised brain MRI anomaly detection using both VAE and DDPM backbones. The primary source was the IXI “healthy” set, with evaluation on FastMRI+ images containing 13 pathologies.

Key quantitative results include:

Pathology	Baseline TPR	AdaptiveAD TPR	TPR Gain
Enlarged Ventricles	0.47	0.84	+78%
Edema	0.72	0.83	+15%
Average (all; RA)	—	—	+37%
Average (all; DDPM)	—	—	+87%

MAE reductions up to 25%, SSIM increases up to 5%, and LPIPS reductions up to 30% were observed on held-out healthy images.

Qualitatively, false positives from scanner-driven intensity artifacts were effectively ablated after adaptation, while pathological boundaries remained sharp. Entropy of the anomaly prediction map was halved, aligning with entropy minimization in test-time learning (Ambekar et al., 4 Oct 2024).

6. Relation to Other AdaptiveAD Formulations

The AdaptiveAD paradigm as described above is distinct from (but partially related to) several other areas employing “adaptive” and “AD” terminology:

• Adaptive NAD: This refers to an Adaptive Network Anomaly Detector for online, self-adaptive, interpretable network anomaly detection using a two-layer pipeline (deep LSTM-VAE plus interpretable RF with adaptive thresholding) for cyber-security in IoT applications (Yuan et al., 30 Oct 2024).
• Adaptive Defense in Active Directory: Stackelberg-game-theoretic frameworks combine GNN-approximated dynamic programming and evolutionary diversity optimization (EDO) for attacker-defender co-evolution, relevant to blocking policy optimization in Active Directory graphs (Goel et al., 16 May 2025, Goel et al., 2022).
• AdaptiveAD in Autonomous Driving: Here, it designates a dual-branch architecture for end-to-end driving models that decouple scene and ego signals, fusing their predictions via an adaptive gating mechanism to enhance generalization and avoid over-reliance on ego-status priors (Tang et al., 17 Nov 2025).
• Adaptive Optimization (ADMM derivatives): Several methods (e.g., Adaptive Stochastic ADMM (Zhao et al., 2013), Adaptive Relaxed ADMM (Xu et al., 2017), Adaptive Consensus ADMM (Xu et al., 2017)) use “AdaptiveAD” as shorthand for adaptive penalty or proximal term selection in distributed or stochastic optimization, but this is functionally and contextually distinct from anomaly detection.

7. Limitations, Extensions, and Outlook

The primary limitation of AdaptiveAD in anomaly detection is the use of simple Gaussian thresholds for masking anomalies in feature space; more sophisticated outlier detection (e.g., Mahalanobis scores, learned maskers) may yield finer-grained adaptation. Current adaptation speeds are limited (minutes per image slice); integration with meta-learning or accelerated optimizers is suggested as a mitigation.

While AdaptiveAD is currently tailored to reconstruction-based anomaly detectors, its model-agnostic, wraparound paradigm makes it extensible to hybrid discriminative-generative methods and other imaging modalities beyond MRI. The approach’s conceptual decoupling of distributional drift from pathological shift is broadly applicable in domain-adaptive medical imaging and could be generalized to structured, sequential, and tabular anomaly detection.

References:

• “Selective Test-Time Adaptation for Unsupervised Anomaly Detection using Neural Implicit Representations” (Ambekar et al., 4 Oct 2024)
• “Adaptive NAD: Online and Self-adaptive Unsupervised Network Anomaly Detector” (Yuan et al., 30 Oct 2024)
• “Decoupling Scene Perception and Ego Status: A Multi-Context Fusion Approach for Enhanced Generalization in End-to-End Autonomous Driving” (Tang et al., 17 Nov 2025)
• “Co-Evolutionary Defence of Active Directory Attack Graphs via GNN-Approximated Dynamic Programming” (Goel et al., 16 May 2025)
• “Adaptive Stochastic Alternating Direction Method of Multipliers” (Zhao et al., 2013)