Extending Unsupervised Meta-Learning with Latent-Space Interpolation in GANs to Semi-Supervised Meta-Learning

📄 Stanford CS330 Final Project
👩‍💻 Helgi Hilmarsson, Davide Giovanardi, Arvind Kumar, William Steenbergen
🎓 Stanford University – ICME
🧑‍🏫 Mentor: Rafael Rafailov

---

📌 Abstract

While supervised and unsupervised meta-learning techniques are widely researched, semi-supervised approaches remain underexplored.
In this work, we extend LASIUM (Latent-Space Interpolation for Unsupervised Meta-Learning, Khodadadeh et al.) into a semi-supervised framework.

We propose two methods:

• SSML (Semi-Supervised Meta-Learning): Concatenates labeled data with GAN-generated data for meta-training.
• SSML-SSG (Semi-Supervised Meta-Learning with Semi-Supervised GAN): Uses labeled data both in meta-training and to improve GAN sample quality.

We evaluate these on Omniglot and Mini-ImageNet, showing that adding labeled data improves performance on unseen tasks monotonically.
Interestingly, SSML outperforms SSML-SSG on Omniglot, possibly because SS-GANs generate noisier samples.

---

📖 Table of Contents

1. Introduction
2. Related Work
3. Methods
   • SSML
   • SSML-SSG
   • Semi-Supervised Data Pipeline
   • Architectural Details
4. Experiments
   • Datasets
   • Experimental Details
5. Results
6. Conclusion
7. Discussion and Next Steps
8. Implementation
9. Contributions

---

🚀 Introduction

Meta-learning algorithms aim to prepare models to quickly adapt to new tasks.
We explore how limited labeled data can enhance unsupervised meta-learning, with applications in domains like medical imaging where labels are scarce.

---

📚 Related Work

• CACTUs (Hsu et al.): Uses clustering to generate pseudo-labels.
• LASIUM (Khodadadeh et al.): Uses GAN/VAEs for latent space interpolation to build supervised tasks from unlabeled data.
• Semi-supervised GANs (Salimans et al.): Extend GAN discriminators with label prediction for better feature learning.

---

🛠 Methods

🔹 Semi-Supervised Meta-Learning (SSML)

• Combines labeled data with GAN-generated tasks.
• The meta-learner alternates updates between real and synthetic data.

🔹 Semi-Supervised Meta-Learning with Semi-Supervised GAN (SSML-SSG)

• Extends GAN discriminator with label prediction.
• Generator learns to produce class-conditional images.

🔹 Semi-Supervised Data Pipeline

We simulate label availability with three schemes:

1. X% of instances labeled across all classes.
2. 100% of instances labeled for X% of classes.
3. Randomly label X% of the dataset.

🔹 Architectural Details

• Omniglot: 3-layer convolutional GAN.
• Mini-ImageNet: 5-block convolutional GAN (more compute-intensive).
• Meta-learning via MAML.

---

🧪 Experiments

Datasets

• Omniglot: Handwritten characters (low intra-class variance).
• Mini-ImageNet: Complex, diverse images (high intra- and inter-class variance).

Setup

• 1-shot, 5-way classification.
• GAN trained for 500 epochs on Omniglot and 100 epochs on Mini-ImageNet.

---

📊 Results

Omniglot (1-shot, 5-way classification)

• Both SSML and SSML-SSG improve over unsupervised LASIUM.
• SSML slightly outperforms SSML-SSG under different label sampling strategies.
• Suggests that adding labeled data consistently improves downstream performance.

Figure 7: Performance on Omniglot using LASIUM-RO. Steep curve shows large performance gains even with limited labels.

---

Mini-ImageNet (1-shot, 5-way classification)

• Both methods show improvements as label percentage increases.
• Performance curve is more gradual than Omniglot, reflecting greater dataset complexity.
• Due to compute limits, only SSML was fully evaluated.

Figure 8: Performance on Mini-ImageNet using LASIUM-N. Gradual improvement curve reflects dataset diversity.

---

GAN vs SS-GAN Outputs

• Regular GAN produces visually richer Omniglot samples.
• SS-GAN samples contain more noise and simpler strokes, possibly explaining lower performance.

---

✅ Conclusion

• Semi-supervised extensions improve over unsupervised LASIUM.
• SSML outperforms SSML-SSG on Omniglot due to better-quality synthetic tasks.
• Provides guidance on how much labeled data is worth annotating before resources are spent.

---

💡 Discussion & Next Steps

• Improve GAN training with architecture optimization and feature matching.
• Explore SS-GAN on Mini-ImageNet, where noise may align better with natural class variance.
• Extend to other meta-learners (e.g., ProtoNets) and other unsupervised methods (e.g., DeepCluster).

---

📌 This work was conducted as part of Stanford CS330: Deep Multi-Task and Meta Learning.