Conditional Neural Processes

CNPs parameterize a distribution over functions rather than a single function. Given a context set of input-output pairs, the model defines a predictive distribution for any target input. This captures epistemic uncertainty—the model knows what it doesn't know—shrinking predictive variance near observed context points and expanding it in unexplored regions. Unlike Gaussian processes, CNPs scale linearly with context size, not cubically.

The CNP consists of two core components:

• Encoder: A permutation-invariant neural network (typically a DeepSet) that aggregates each context point into a fixed-dimensional latent representation r. Permutation invariance ensures the model treats context sets, not sequences.
• Decoder: A feedforward network that takes the aggregated representation r and a target input x* to output the parameters of a Gaussian predictive distribution: mean μ and variance σ².

This split design enforces the conditional independence of target predictions given the context summary.

A defining property of CNPs is that the predictive distribution p(y*|x*, C) is invariant to the ordering of context points in C. This is achieved by the encoder aggregating per-point representations using a commutative operation:

• Mean aggregation: r = (1/|C|) Σ h(x_i, y_i)
• Sum aggregation: r = Σ h(x_i, y_i)
• Self-attention: Used in Attentive CNPs to weight context points by relevance to the target.

This property is essential for meta-learning, where the support set has no inherent order.

CNPs are trained via episodic meta-learning. During training, the model sees many tasks (functions) sampled from a distribution. For each task, a random subset of points forms the context set, and the remaining points form the target set. The loss maximizes the log-likelihood of the target points given the context. At test time, the model can make predictions for a completely new function after observing only a handful of context points—no gradient updates required.

In signal classification, CNPs treat each novel modulation type as a new function to model. The context set consists of a few labeled IQ samples, and the target set contains unlabeled samples to classify. Key advantages:

• Uncertainty calibration: Provides confidence scores for rejecting unknown modulations
• Sample efficiency: Learns to classify from 1-5 examples per class
• Continuous adaptation: Naturally handles an open-ended set of modulation formats without retraining
• Channel robustness: Can condition on channel parameters as additional context

CNPs learn to model a distribution over functions rather than a single deterministic mapping. In automatic modulation classification, this means the model captures the statistical regularities across hundreds of base modulation tasks. When presented with a small support set of IQ samples from a novel modulation scheme, the CNP conditions on these context points to infer a posterior predictive distribution over the query signal's modulation type. This provides not just a classification but a well-calibrated uncertainty estimate, critical for spectrum monitoring systems that must flag ambiguous or unknown signals for human analysis.

Unlike fixed-architecture few-shot learners that require a rigid N-way K-shot setup, CNPs can condition on an arbitrary number of context points at inference time. For a SIGINT operator intercepting a burst transmission, the model can ingest all available clean IQ samples—whether 3 or 30—without architectural modification. The permutation-invariant encoder aggregates context observations via a symmetric function (e.g., mean pooling), ensuring the representation is independent of the order in which support samples are presented. This flexibility is essential for dynamic spectrum environments where the number of observable signal exemplars varies unpredictably.

CNPs achieve few-shot modulation recognition through amortized inference, not iterative optimization. During meta-training, the model learns a shared encoder-decoder architecture that implicitly captures how to extract task-specific representations from context sets. At deployment, classifying a new modulation type requires only a single forward pass through the network—no fine-tuning, no gradient steps, no weight updates. This contrasts sharply with optimization-based meta-learners like MAML, which require multiple inner-loop gradient steps. For real-time spectrum classification on edge hardware, this zero-gradient adaptation translates directly to microsecond-level inference latency.

A critical limitation of deterministic few-shot classifiers is their inability to express epistemic uncertainty—they may confidently misclassify a completely unknown modulation scheme. CNPs inherently output a predictive distribution with a mean and variance for each query point. In open set signal recognition, the variance term serves as a natural anomaly score:

• Low variance: The context set strongly constrains the prediction; the signal is likely a known class.
• High variance: The context provides insufficient information; the signal may be novel or adversarial. This principled uncertainty quantification enables autonomous spectrum monitoring systems to trigger alerts and queue signals for expert review.

Standard CNPs suffer from underfitting because the independent Gaussian predictive distribution at each query point fails to capture correlations between nearby frequency or time bins. Attentive Conditional Neural Processes (AttnCNPs) address this by incorporating a multi-head cross-attention mechanism between the context set and each query location. For modulation classification, this allows the model to learn that IQ samples at adjacent symbol periods share statistical structure. The attention weights explicitly reveal which context points the model relies on for each classification decision, providing a form of interpretability valuable for electronic warfare analysts who need to audit automated decisions.

Training CNPs for modulation recognition requires diverse episodic tasks spanning many modulation families. For rare or classified signal types where real over-the-air captures are scarce, synthetic context sets can be generated using signal simulation toolchains (e.g., GNU Radio, MATLAB). During meta-training, each episode constructs a support set by sampling from a synthetic modulation distribution with randomized channel impairments—fading, frequency offset, phase noise. The CNP learns to extract channel-invariant features that generalize from synthetic context to real query signals. This sim-to-real transfer capability dramatically reduces the burden of collecting labeled RF data for every modulation scheme of interest.