A focused exploration into enhancing the accuracy of MNIST digit predictions through semi-supervised learning with a minimal set of 100 labeled samples. This project employs pseudo-labels to effectively use both labeled and unlabeled data, utilizing TensorFlow to build and implement models. It serves as a concise demonstration of improving prediction accuracy with limited labeled data.
The methodology adopts a two-phase approach, emphasizing the practical application of semi-supervised learning algorithms enhanced by pseudo-labels:
- Baseline Model Development: Constructs a foundational model using neural network techniques to address the MNIST dataset.
- Semi-Supervised Learning with Pseudo-Labels: Augments model performance through the generation and use of pseudo-labels, capitalizing on the unlabeled data within the dataset for more effective learning.
- MNIST Dataset: The primary dataset comprising 70,000 black-and-white images of handwritten digits, divided into training and validation sets.
- Neural Networks (NN) and Convolutional Neural Networks (CNN): These form the basis of our models, further refined through semi-supervised learning strategies.
- Semi-Supervised Self Training: This innovative technique utilizes pseudo-labels for training, aiming to leverage the unlabeled dataset effectively.
Detailing the project's implementation, this section outlines the data preparation, architectural decisions for NN and CNN models, and the integration of dropout and pseudo-labels to prevent overfitting and boost model generalization. Utilizing TensorFlow and Keras, the project showcases a structured approach to model development and training.