Mode Collapse

The generator maps multiple distinct latent space points to identical or near-identical outputs. Instead of sampling from the full target distribution, the generator collapses to producing a single mode or a small family of highly similar RF waveforms. In signal generation tasks, this means the synthetic dataset may contain only QPSK variants at a single SNR level, completely omitting 16QAM, 64QAM, or other modulation classes. The Inception Score and Fréchet Inception Distance (FID) metrics will show severe degradation as intra-class diversity vanishes.

Rather than stabilizing on a single mode, the generator may cycle between a small set of modes over training iterations. The discriminator adapts to reject one collapsed output, causing the generator to switch to a different narrow mode that temporarily fools it. This cat-and-mouse dynamic never converges to a diverse equilibrium. In RF applications, the generator might alternate between producing only BPSK and only 16QAM signals, never learning to produce both simultaneously or to cover intermediate modulation orders.

When the generator collapses, the discriminator quickly learns to reject the limited set of fake samples with near-perfect accuracy. The discriminator's loss approaches zero, and its gradient signal to the generator vanishes—a phenomenon known as vanishing gradients. Without meaningful feedback, the generator cannot recover diversity. In RF GANs, a discriminator trained on rich real-world spectrum data will trivially distinguish a collapsed generator's repetitive outputs, halting all useful learning.

Mode collapse disproportionately affects rare or underrepresented signal classes. The generator optimizes for the modes that most easily fool the discriminator, typically the majority classes in the training distribution. In RF signal intelligence datasets with severe class imbalance—where common modulations vastly outnumber exotic ones—mode collapse causes the complete omission of rare but operationally critical signal types such as spread-spectrum or frequency-hopping waveforms from the synthetic training set.

The Wasserstein GAN (WGAN) architecture replaces the standard binary cross-entropy loss with the Earth-Mover's distance, providing a smoother gradient landscape that resists collapse. The WGAN critic outputs a real-valued score rather than a probability, and gradient penalty enforcement on the critic ensures Lipschitz continuity. In RF waveform generation, WGAN-GP variants consistently produce more diverse synthetic constellations and power spectral densities compared to vanilla GANs, though at increased computational cost.

Mode collapse can be empirically detected by analyzing the pairwise similarity of generated samples. Techniques include:

• Computing the mean squared error (MSE) between batches of generated IQ samples—collapse produces abnormally low variance
• Visualizing the t-SNE or UMAP embeddings of generated vs. real signals to identify clustering into few tight groups
• Measuring the number of distinct modulation constellations produced when conditioning on different class labels
• Tracking the effective rank of the generator's output covariance matrix over training iterations

A generative architecture that feeds auxiliary information—such as modulation type, SNR level, or device label—into both the generator and discriminator during training. This conditioning forces the model to associate specific data modes with explicit labels.

• Prevents the generator from ignoring minority classes by making the discriminator aware of class context
• Enables controlled synthesis: generate only QPSK signals at 10dB SNR on demand
• The discriminator evaluates both realism and class consistency, penalizing mode-dropping behavior
• Essential for RF data augmentation where balanced class representation is critical for downstream classifiers

A training stabilization technique that dynamically applies a pipeline of augmentations to both real and generated samples before they enter the discriminator. By preventing the discriminator from memorizing the limited training set, ADA reduces overfitting that triggers mode collapse.

• Augmentations include additive Gaussian noise, frequency shifting, and time warping tailored for RF signals
• The augmentation probability is adaptively controlled based on discriminator overfitting heuristics
• Particularly effective in limited RF data regimes where the discriminator rapidly overpowers the generator
• Preserves signal semantics while forcing the discriminator to learn invariant features

A training strategy where the generator's loss function incorporates the discriminator's response to future generator states. Instead of optimizing against the current discriminator, the generator unrolls K steps of discriminator optimization and minimizes the loss at the unrolled state.

• Prevents the generator from exploiting transient discriminator weaknesses that lead to mode collapse
• The generator learns to produce samples that remain convincing even after the discriminator adapts
• Computationally expensive due to the inner optimization loop, but highly effective for stabilizing training
• Useful for complex RF distributions with many overlapping signal modes in high-dimensional IQ space

A discriminator architecture modification that allows the model to evaluate diversity within a minibatch of generated samples. The discriminator computes pairwise similarity statistics between samples and uses these as additional features for classification.

• If the generator produces near-identical samples, the discriminator detects the lack of variation and penalizes it
• The technique explicitly encodes the intuition that a batch of real data exhibits higher internal diversity
• Implemented by projecting sample features into a learned space and computing L1 distances between all pairs
• Provides a direct gradient signal pushing the generator away from collapsed modes toward full distribution coverage

A simple yet effective modification where the discriminator processes packed sets of samples—concatenating multiple real or generated examples into a single input—rather than evaluating individual samples. This forces the discriminator to assess joint distributional properties.

• The discriminator learns to detect when a packed set lacks diversity, directly penalizing mode collapse
• Requires minimal architectural changes: simply stack samples along the batch or channel dimension
• The generator must produce varied outputs within each pack to fool the discriminator
• In RF signal generation, packing multiple IQ sequences together enables the discriminator to evaluate cross-signal diversity