Data-Efficient Instance Generation from Instance Discrimination (2106.04566v1)

Abstract: Generative Adversarial Networks (GANs) have significantly advanced image synthesis, however, the synthesis quality drops significantly given a limited amount of training data. To improve the data efficiency of GAN training, prior work typically employs data augmentation to mitigate the overfitting of the discriminator yet still learn the discriminator with a bi-classification (i.e., real vs. fake) task. In this work, we propose a data-efficient Instance Generation (InsGen) method based on instance discrimination. Concretely, besides differentiating the real domain from the fake domain, the discriminator is required to distinguish every individual image, no matter it comes from the training set or from the generator. In this way, the discriminator can benefit from the infinite synthesized samples for training, alleviating the overfitting problem caused by insufficient training data. A noise perturbation strategy is further introduced to improve its discriminative power. Meanwhile, the learned instance discrimination capability from the discriminator is in turn exploited to encourage the generator for diverse generation. Extensive experiments demonstrate the effectiveness of our method on a variety of datasets and training settings. Noticeably, on the setting of 2K training images from the FFHQ dataset, we outperform the state-of-the-art approach with 23.5% FID improvement.

An Analysis of Data-Efficient Instance Generation from Instance Discrimination

The paper "Data-Efficient Instance Generation from Instance Discrimination" addresses key challenges associated with Generative Adversarial Networks (GANs), particularly the drop in synthesis quality when training data is limited. Typically, GANs require substantial datasets to train successfully, as the discriminator in a GAN model tends to overfit on smaller datasets, hampering the quality of image synthesis by the generator. The authors propose a novel method called Instance Generation (InsGen) which introduces an instance discrimination task during GAN training to enhance data efficiency.

Methodology

The core premise of the InsGen method revolves around modifying the training objective of the GAN discriminator to include instance discrimination. Conventionally, a discriminator distinguishes between real and fake data, but InsGen requires the discriminator to recognize each image instance individually, whether it originates from the real dataset or is generated by the generator. This approach aims to harness the synthetic data effectively to augment the discriminator's training, thereby addressing overfitting issues. Key innovations of the proposed approach include:

1. Instance Discrimination: Tasking the discriminator with a more complex classification challenge than the binary real versus fake task. This is achieved by considering each input image as a separate classification category.
2. Noise Perturbation Strategy: This strategy introduces minor perturbations in the latent space for fake images, enhancing the difficulty of the discrimination task and consequently bolstering the discriminator's power.
3. Reciprocal Learning Objective: The capabilities learned by the discriminator through instance discrimination are looped back to encourage the generator towards diverse generation. This adjustment ensures that outputs from the generator are not only realistic but also varied and distinguishable from each other.
4. Data Augmentation: The approach utilizes synthesized data effectively as a complementary data source, granting the discriminator access to an "infinite" data pool and thereby boosting its capability without requiring real, labeled data.

Experimental Results

The effectiveness of the InsGen method is validated through extensive experiments on various datasets. Specifically, the paper illustrates significant improvements in image generation quality measured by the Fréchet Inception Distance (FID). For example:

• On a dataset with only 2,000 training images from the FFHQ dataset, the proposed InsGen approach yields an FID score improvement of 23.5% over previous state-of-the-art methods.
• This performance gain is consistent across other data settings, such as datasets with 10,000 and 70,000 images, demonstrating the robustness and versatility of the method.

Implications and Future Work

The introduction of InsGen marks a notable advancement in data-efficient GAN training methodologies. By enhancing the discriminative power of the discriminator, the generator can be trained to achieve higher quality and more diverse image synthesis with limited data. This approach could be extrapolated to applications where extensive datasets are infeasible, offering pathways for applying GANs in areas with privacy constraints or sparse data availability.

Looking forward, future research could explore further innovations in instance discrimination tasks and their potential synergies with other self-supervised learning strategies. Additionally, investigating the scalability of this framework to different image domains or GAN architectures could provide broader insights into its applicability and generalizability to various generative tasks or network configurations.

Conclusion

The InsGen framework presents an innovative modification to the conventional GAN training regime, enabling high-quality synthesis under limited data conditions. As the landscape of generative models continues to evolve, methods like InsGen that address data efficiency head-on represent important milestones towards more broadly applicable and resource-conscious AI solutions.