Aspect-Distributed Prototype Strategy

• ADP is a meta-learning approach that builds multiple class prototypes over distinct aspect regions to address intra-class variability.
• It partitions the aspect domain and employs pre-trained embeddings to reduce prototype confusion in few-shot learning scenarios.
• Empirical results in radar ATR and DFOS demonstrate improved accuracy, with gains up to 16.7% over single-prototype methods.

Aspect-Distributed Prototype (ADP) Strategy is a meta-learning approach developed to enhance robustness and generalization in classification tasks under few-shot regimes where the data exhibits significant aspect-induced intra-class variance. The ADP strategy centers around constructing multiple class prototypes distributed over contiguous aspect regions, capturing the underlying multi-modal nature of samples that arise due to physical, environmental, or sensor viewpoint variations. This allows models, including LLMs for radar Automatic Target Recognition (ATR) and dual-domain learners for distributed fiber optic sensing (DFOS), to mitigate overfitting and prototype confusion commonly observed in small support set scenarios. ADP has shown empirical superiority over monolithic (single-prototype) approaches in both simulated and measured datasets for HRRP ATR and DFOS activity identification (Bi et al., 7 Dec 2025, He et al., 22 Nov 2025).

1. Motivation and Problem Context

The ADP strategy addresses classification challenges where input data is sensitive to "aspect"—a general term encompassing physical angle, sensor domain, viewpoint, or environmental context. In HRRP ATR, aspect refers specifically to radar azimuth and pitch angles, which induce significant variability in observable scattering centers (SCs). For DFOS, aspects correspond to the temporal (waveform) and frequency (spectrogram) domains, with further intra-class diversity caused by fiber deployment and environmental factors.

Existing meta-learning and prototype-based approaches such as HRRPLLM for radar ATR merge all support samples into a single class prototype, failing to account for aspect-driven sample diversity. This leads to prototype confusion—where increased support set size or aspect span can degrade recognition accuracy, especially when testing samples differ in aspect from those seen in support (Bi et al., 7 Dec 2025). In DFOS settings, domain shift between deployment types and signal modalities similarly causes conventional mean-prototype methods to generalize poorly (He et al., 22 Nov 2025).

2. ADP Framework and Formalism

ADP generalizes the concept of prototypes by introducing aspect-wise class representation. Let $C$ denote the set of classes, $K$ the number of support samples per class, and $\Theta$ the aspect domain (e.g., angular space, data view space). The key steps are:

1. Partitioning by Aspect: Divide $\Theta$ into $A$ disjoint regions $\{\Theta_a\}$ (e.g., equal-width bins for angle, or separate temporal/frequency views).
2. Aspect-wise Support Sets: For each class $c$ and aspect region $a$, construct $$S_{c,a} = \{(x_i,\theta_i) \in DS\,|\,y_i=c~\text{and}~\theta_i\in\Theta_a\}$$, where $DS$ is the support set.
3. Prototype Computation: Use a pre-trained embedding function $f:\mathcal{X}\rightarrow \mathbb{R}^d$ (radar or fiber input to feature space) to compute aspect-distributed prototypes:

$$\mu_{c,a} = \frac{1}{|S_{c,a}|}\sum_{(x,\theta)\in S_{c,a}} f(x)$$

For dual-domain learning, embeddings for each domain are clustered to multiple prototypes per aspect (He et al., 22 Nov 2025).

1. Prototype Aggregation: Optionally average across aspect regions for a single prototype, though ADP typically retains the set $\{\mu_{c,a}\}$ for downstream classification.

This formalism captures intra-class variation and provides explicit structure for meta-learners to select the most appropriate prototype during query matching.

3. Query Matching and Classification Rules

ADP modifies the traditional classification process by allowing queries to select from aspect-distributed prototypes:

1. Distance Metric: For each query $x$ (e.g., HRRP waveform, DFOS sample), compute the distance to prototypes in each aspect region:

$d(f(x), \mu_{c,a}) = \|f(x) - \mu_{c,a}\|_2$

For dual-domain DFOS learners, similarities (e.g., cosine) are weighted by domain guidance and prototype-specific sensitivity parameters (He et al., 22 Nov 2025).

2. Minimum-over-Aspects: Assign the class score as the negative minimum distance across all aspect prototypes:

$s_c(x) = -\min_{a} d(f(x), \mu_{c,a})$

3. Softmax Probability: Convert scores to class probabilities:

$$p(c|x) = \frac{\exp(s_c(x))}{\sum_{c'}\exp(s_{c'}(x))}$$

4. Query-Adaptive Aggregation (DFOS): Attention and guidance weights $\alpha_a, \beta_{a,i}$ modulate multi-prototype aggregation for each domain/aspect, which are combined through a relation network $f_{\rm rel}$ to form the final decision logits (He et al., 22 Nov 2025).

This approach allows for dynamic matching of queries to the prototype most representative of their aspect, markedly improving robustness to mismatched support and query aspects.

4. In-Context Learning and Statistical Guidance

ADP integrates with both frozen embedding backbones and LLM-centric in-context learning for radar ATR, as well as statistical guidance networks in DFOS:

• Radar ATR with LLMs: ADP-based prompts enumerate aspect-clustered prototypes with associated SC lists. The LLM (e.g., GPT-4.1-ADP) receives a task description and candidate prototypes, then predicts the target class based on explicit reasoning over aspect-distributed SCs. No fine-tuning of LLM weights is employed—only prompt engineering leverages ADP structure (Bi et al., 7 Dec 2025).
• Statistical Guidance Network (SGN, DFOS): SGN computes global input statistics, generating domain importance weights $\alpha_t, \alpha_f$ (temporal and frequency), prototype sensitivity scalars $\beta_{a,i}$, and a guidance vector $\mathbf{g}$. These parameters modulate the prototype attention mechanism during query matching, providing a data-driven prior for adaptive aggregation (He et al., 22 Nov 2025).

Training employs cross-entropy over queries and regularization to encourage diversity among aspect prototypes. All embedding and SGN parameters are typically learned via episode-wise gradient descent.

5. Empirical Results and Robustness

ADP has undergone validation in distinct application domains, demonstrating superior generalization under aspect variance and data scarcity:

Task	ADP Variant	Dataset	Accuracy Gain	F1 Gain
HRRP ATR (Simulated, 10-shot)	GPT-04-mini-ADP	12 aircraft	+16.7% over HRRPLLM	+—
HRRP ATR (Measured, 20-shot)	GPT-4.1-ADP	3 aircraft	+3.34% over HRRPLLM	+5.76%
DFOS OSDG1	DUPLE (ADP inside)	DFOS data	+8% over vanilla prototype	+—
DFOS OSDG2	DUPLE (ADP inside)	DFOS data	+12% over vanilla prototype	+—

ADP alleviates the trend of declining accuracy with increased support set size observed in monolithic prototype approaches under aspect heterogeneity. For DFOS, the combination of dual-domain multi-prototype learning, statistical guidance, and query-aware aggregation achieves a total performance improvement of approximately 12–15% in accuracy and F1, while avoiding catastrophic collapse on minority classes. This suggests that ADP is robust to both domain shift and intra-class diversity in scenarios where labeled data is limited (Bi et al., 7 Dec 2025, He et al., 22 Nov 2025).

6. Comparative Approaches and Applications

ADP subsumes several previous prototype-based and meta-learning methods but introduces key innovations:

• Monolithic Prototype Methods: Suffer from overfitting and poor generalization when query aspect differs from the support set (Bi et al., 7 Dec 2025).
• Dual-Domain Multi-Prototype (DFOS): Fuses temporal and frequency embeddings, producing multiple aspect-aware prototypes per class (He et al., 22 Nov 2025).
• Statistical Guidance and Query-Aware Aggregation: SGN interprets global statistics to dynamically adapt prototype relevance, providing resilience to domain shifts and environmental changes.

ADP is directly applicable to HRRP ATR, DFOS, and other domains exhibiting strong aspect-dependent sample variability. It operationalizes robust few-shot learning by structuring intra-class prototype representations and leveraging in-context or guidance-driven reasoning for classification.

7. Directions and Implications

The success of ADP in mitigating aspect sensitivity and enhancing performance in low-data, multi-modal environments indicates its utility for a broad range of sensing and recognition tasks. A plausible implication is the potential extension of ADP strategies to multi-sensor fusion, cross-view generalization, and domains with complex attribute-driven heterogeneity. Incorporating statistical guidance and query-adaptive weighting mechanisms may further bolster generalization to real-world applications where environmental and domain factors can fluctuate unpredictably. Empirical trends suggest that maintaining aspect-distributed prototypes and integrating attention-guided reasoning may become foundational principles in future meta-learning and few-shot ATR frameworks (Bi et al., 7 Dec 2025, He et al., 22 Nov 2025).