Introduction
Training conditional generative models on multi-source data presents significant challenges, primarily due to domain shift and the "curse of dimensionality." When data is pooled from disparate sources—such as different hospitals or financial institutions—"outlier" domains can lead to negative transfer, where the inclusion of more data actually degrades model performance. Furthermore, standard Generative Adversarial Networks (GANs) struggle to capture high-dimensional distributions efficiently without massive pre-training.
This paper introduces STWGAN (Selected Transfer-WGAN), a framework designed to address these issues by leveraging low-dimensional invariant representations and an automated source selection mechanism. STWGAN filters out unreliable data sources and focuses on shared underlying structures, achieving faster convergence and higher accuracy in niche domains where large-scale pre-trained models are unavailable.
Methodology: The STWGAN Framework
The STWGAN approach operates in two distinct stages to ensure both robustness to outliers and statistical efficiency.
Stage 1: Representation and Selection
The model first learns a mapping $R(X)$ that compresses high-dimensional inputs $X$ into a low-dimensional, domain-invariant representation $Z$. To mitigate negative transfer, the framework employs a selection criterion based on the 1-Wasserstein distance. By comparing the distribution of source domains against the target domain in this latent space, the model identifies and discards "distant" outlier sources.
Stage 2: Refined Generative Learning
Once reliable sources are identified, the final conditional generator is trained using a combination of target data and the selected "reliable" source data. Unlike traditional transfer learning that focuses on parameter fine-tuning, STWGAN transfers knowledge through these shared representations. Theoretically, this reduces the convergence rate from $n^{-1/(d+q)}$ to $n^{-1/(r+q)}$, where $r$ is the reduced dimension of the latent space, significantly improving learning efficiency.
Experimental Evaluation
The researchers validated STWGAN across several benchmarks:
- Numerical Simulations: Testing multimodal distribution capture in nonlinear and mixture models.
- MNIST Digit Reconstruction: Predicting missing portions of handwritten digits.
- Image-to-Image Translation: Utilizing the Edges2Shoes and Edges2Handbags datasets.
- Tabular Data: Predicting biological markers in the Abalone dataset.
Results and Discussion
STWGAN consistently outperformed "Target-Only" and "Naive Pooling" baselines across all metrics. In numerical simulations, it achieved lower Mean Squared Error (MSE) for conditional mean and variance.
In image-to-image tasks, as shown in the figure above, STWGAN produced significantly more realistic and vibrant results compared to baselines, which often generated dull or inconsistent patterns. On the MNIST dataset, the model demonstrated superior "stroke consistency," effectively using source knowledge to disambiguate similar digits (e.g., distinguishing an '8' from a '6' given partial information).
The results highlight two key insights:
- Quality Over Quantity: Selective data pooling prevents the performance degradation associated with negative transfer.
- Dimensionality Reduction: Operating in a lower-dimensional latent space fundamentally improves the statistical efficiency of the generative process, allowing the model to "borrow strength" from sources even with limited target samples.
Conclusion and Future Work
STWGAN provides a robust and efficient solution for multi-source conditional generative learning, particularly in data-scarce or niche environments. Future research directions include exploring the internal mechanisms by which neural networks select invariant structures and integrating these selection techniques into state-of-the-art Diffusion Models. Additionally, refining the selection threshold ($M$) remains a practical priority for balancing robustness and data utility.
Summary updated