t-Distributed Stochastic Neighbor Embedding (t-SNE)

t-SNE converts high-dimensional Euclidean distances into conditional probabilities that represent similarities. It constructs a probability distribution over pairs of high-dimensional objects such that similar objects have a high probability of being picked, while dissimilar points have an extremely low probability. A similar distribution is constructed in the low-dimensional map, and the algorithm minimizes the Kullback-Leibler divergence between the two distributions using gradient descent. This probabilistic framework is what allows t-SNE to faithfully represent the neighborhood structure of complex signal feature embeddings.

Unlike Gaussian kernels, t-SNE employs a Student-t distribution with one degree of freedom in the low-dimensional space. This heavy-tailed distribution is the defining innovation that solves the 'crowding problem' inherent in earlier methods like SNE. The t-distribution allows moderately dissimilar points in the high-dimensional space to be modeled by much larger distances in the low-dimensional map, preventing unwanted attractive forces between dissimilar clusters. This creates more distinct, well-separated visual clusters when plotting emitter signatures.

Perplexity is a crucial hyperparameter that balances attention between local and global aspects of the data. It can be interpreted as a smooth measure of the effective number of neighbors and is typically set between 5 and 50. In RF fingerprinting:

• Low perplexity (5-10): Reveals fine-grained local clusters of nearly identical hardware
• High perplexity (30-50): Captures broader manufacturer or model-level groupings Performance is generally robust across a range, but the choice significantly impacts the visual separation of emitter classes.

The t-SNE cost function is non-convex, meaning gradient descent converges to a local minimum that depends on the random initialization. This stochastic nature produces different visual outputs on each run. Key implications for signal analysis:

• Reproducibility: Always set a fixed random seed for consistent results
• Topology vs. Distance: The relative sizes of clusters and distances between them are not interpretable; only the grouping structure matters
• Multiple Runs: Running t-SNE several times with different initializations can confirm the stability of observed emitter clusters

While powerful for visualization, t-SNE has critical limitations in production RF systems:

• No parametric mapping: It does not learn a function to map new, unseen signals into the embedding; the entire dataset must be re-run
• Global structure distortion: Distances between well-separated clusters are meaningless; t-SNE only guarantees local neighborhood fidelity
• Scalability bottleneck: The original Barnes-Hut implementation is O(N log N), but standard t-SNE is O(N²), making it impractical for massive real-time signal streams For these reasons, t-SNE is used for exploratory analysis, while UMAP or parametric methods are preferred for production deployment.

In deep learning signal identification, t-SNE is applied to the penultimate layer activations of a neural network trained on IQ data or spectrograms. By projecting these high-dimensional feature embeddings to 2D, engineers can visually validate:

• Cluster purity: Whether signals from the same physical transmitter form tight, distinct groups
• Inter-class separation: The degree of overlap between different device fingerprints
• Anomaly detection: Outliers that may represent spoofing attempts or faulty hardware This qualitative assessment guides architecture decisions before deploying models into production.

A manifold learning technique that preserves both local and global data structure better than t-SNE. Unlike t-SNE's focus on local neighborhoods, UMAP uses topological data analysis and Riemannian geometry to maintain meaningful inter-cluster distances. In RF fingerprinting, UMAP excels at revealing hierarchical relationships between different transmitter models and manufacturers.

• Faster runtime than t-SNE on large datasets
• Preserves global structure alongside local clusters
• Scales to millions of emitter samples
• Deterministic output with fixed random seed

The process of mapping high-dimensional IQ samples or spectrograms into a lower-dimensional vector space where semantically similar device signatures cluster together. Before applying t-SNE, a neural network typically extracts a compact embedding vector (e.g., 128-512 dimensions) from raw signal data. The quality of this embedding directly determines whether t-SNE reveals clean, separable device clusters.

• Input: Raw IQ data or time-frequency representations
• Output: Dense vector capturing device-specific features
• Quality measured by intra-class compactness and inter-class separation

The compressed, abstract representation of input data learned by a neural network's bottleneck layer. In RF fingerprinting, the latent space encodes the intrinsic factors of variation related to hardware impairments—oscillator phase noise, I/Q imbalance, power amplifier non-linearity—while discarding irrelevant information like payload content. t-SNE projects this latent space into 2D for visual validation that the model has disentangled device identity from channel effects.

• Dimensionality typically 64-256 before t-SNE projection
• Should exhibit clustering by device, not by channel condition
• Quality assessed via silhouette score and Davies-Bouldin index

A self-supervised paradigm that trains models to produce embeddings where similar signal samples are pulled together and dissimilar ones pushed apart. This directly optimizes the embedding space for t-SNE visualization. Techniques like SimCLR and SupCon use data augmentation—adding channel impairments, noise, and frequency offsets—to create positive pairs from the same transmitter.

• Eliminates need for large labeled datasets
• Produces tightly clustered device representations
• Robust to channel variation when augmentations are well-designed
• t-SNE visualizations reveal cleaner separation than supervised baselines

A metric learning loss function that trains networks to minimize distance between an anchor sample and a positive sample (same device) while maximizing distance to a negative sample (different device). The resulting embeddings form the input to t-SNE, where the margin parameter controls cluster separation. Critical for open-set emitter recognition where unknown devices must fall outside known clusters.

• Loss = max(0, d(anchor, positive) - d(anchor, negative) + margin)
• Hard negative mining essential for discriminative features
• Embeddings directly visualizable with t-SNE for margin tuning

A twin neural network architecture where two identical subnetworks process pairs of signal samples and a distance metric determines if they originated from the same transmitter. The shared-weight design ensures consistent embedding extraction. t-SNE visualizations of Siamese embeddings reveal whether the network has learned a meaningful similarity metric or collapsed to trivial solutions.

• Shared weights guarantee symmetric feature extraction
• Contrastive loss or triplet loss for training
• Ideal for few-shot device enrollment scenarios
• t-SNE used to validate embedding quality before deployment