Omniglot-MetaLearn

This project implements and evaluates two meta-learning approaches on the Omniglot dataset for few-shot image classification:

1. Black-box Meta-Learning (BBML)
2. Model-Agnostic Meta-Learning (MAML)

The objective is to compare their ability to quickly adapt to new character classes using limited training samples.

---

🧠 Meta-Learning Overview

Meta-learning or "learning to learn" trains a model over a distribution of tasks such that it can generalize and adapt quickly to new, unseen tasks with minimal data.

Implemented Approaches:

1. Black-box Meta-Learning (BBML)

• Trains a recurrent or feedforward neural network end-to-end over tasks.
• Uses task conditioning (via support sets) to produce rapid generalization without explicit gradient steps.
• Treats the entire learning process as a black box.

2. Model-Agnostic Meta-Learning (MAML)

• Learns an initialization of model parameters that can be quickly fine-tuned with a few gradient steps on a new task.
• First-order or second-order gradient methods are used.

---

📂 Dataset

• Omniglot (link: Omniglot Dataset)

  • Train: images_background (for meta-training)
  • Test: images_evaluation (for meta-testing)

• Preprocessed using torchvision and split into multiple few-shot learning tasks (e.g., 5-way 1-shot).

---

🧪 Evaluation Setup

• N-way K-shot classification setting (e.g., 5-way 1-shot).
• Performance measured by accuracy on query samples from test tasks.
• Evaluated both methods on unseen classes to measure generalization.

---

📈 Results Summary

| Method | Few-Shot Setup | Accuracy (Test Tasks) | |--------------|----------------|------------------------| | BBML | 5-way 1-shot | ~80.94% | | MAML | 5-way 1-shot | ~85.1% |

• MAML showed better adaptation on unseen tasks with fewer gradient steps.
• BBML performed well but was more sensitive to architecture and hyperparameters.

---

📦 Requirements

Install the dependencies:

Libraries used:

• torch

• torchvision

• numpy

• matplotlib

• tqdm

---

📄 License

MIT License

🙋‍♀️ Acknowledgments

• Based on techniques and code structure from meta-learning tutorials (e.g., MAML paper and GitHub examples).

• Omniglot dataset by Brenden Lake et al.

---