DeepCluster

DeepCluster alternates between two steps: clustering the output features of a ConvNet using k-means and using the resulting cluster assignments as pseudo-labels to train the network via standard supervised cross-entropy loss. This iterative process forces the network to learn semantically meaningful representations without any human annotation. In the RF domain, this enables the automatic grouping of IQ samples by modulation scheme, symbol rate, or transmitter hardware characteristics.

When applied to raw IQ streams, DeepCluster can autonomously partition signals into distinct groups corresponding to modulation types (QPSK, 16QAM, GMSK) without prior knowledge of the modulation scheme. The learned clusters often align with:

• Modulation families (PSK vs. QAM vs. FSK)
• Symbol rates and bandwidth configurations
• Pulse-shaping filter characteristics This capability is critical for spectrum monitoring and electronic support measures where emitter parameters are unknown a priori.

DeepCluster typically employs a standard convolutional architecture such as AlexNet or VGG as the feature extractor, processing spectrograms or raw IQ sequences. The network learns hierarchical representations: early layers capture low-level signal features (carrier frequency, bandwidth), while deeper layers encode higher-level abstractions (modulation type, symbol constellation geometry). The architecture is agnostic to the specific backbone, allowing substitution with ResNet or EfficientNet variants optimized for RF data.

A critical implementation detail is the periodic reassignment of clusters. After each epoch, features are re-extracted and k-means is rerun, preventing the network from overfitting to stale pseudo-labels. However, DeepCluster is susceptible to trivial solutions where all samples map to a single cluster. Mitigation strategies include:

• Uniform sampling across clusters during training
• Variance regularization to prevent feature collapse
• Careful initialization of the k-means centroids using PCA or random sampling

After self-supervised pre-training on unlabeled IQ data, the learned representations can be transferred to downstream tasks with limited labeled samples:

• Few-shot modulation classification: Fine-tune with 5-10 labeled examples per class
• Specific emitter identification: Use cluster centroids as initial prototypes for prototypical networks
• Anomaly detection: Identify signals that fall far from any established cluster centroid The pre-trained backbone serves as a powerful feature extractor, significantly outperforming random initialization when labeled data is scarce.

DeepCluster's reliance on full-dataset k-means clustering at each epoch introduces computational overhead, scaling as O(NKD) where N is dataset size, K is cluster count, and D is feature dimension. For large-scale RF datasets with millions of IQ samples, practical optimizations include:

• Mini-batch k-means for approximate clustering
• Feature caching to avoid redundant forward passes
• Dimensionality reduction via PCA before clustering
• Hierarchical clustering for very large K values (1000+ clusters)

The core unsupervised algorithm at the heart of DeepCluster. It partitions raw IQ embeddings into K distinct groups by iteratively minimizing the within-cluster sum of squares. In the RF domain, this groups signal bursts with similar spectral shapes or hardware impairments without prior labels. The quality of the resulting pseudo-labels is highly dependent on the initialization of centroids and the assumption of spherical clusters, which may not hold for complex modulation manifolds.

A critical failure mode in self-supervised learning where the encoder produces a constant, non-informative vector for all inputs. DeepCluster avoids this trivially because the k-means step forces a partition, but the model can still collapse if all features map to a single cluster. Modern methods like VICReg and Barlow Twins explicitly prevent this through variance and covariance regularization on the embedding batch.

The downstream task that validates DeepCluster's pre-training efficacy. After unsupervised feature learning, the frozen or fine-tuned encoder must classify modulation schemes (BPSK, QAM64, etc.) from only 1-5 labeled examples per class. Prototypical networks compute a class centroid from the support set and classify queries by nearest neighbor in the embedding space, directly leveraging the cluster-friendly features DeepCluster produces.