Skip to content

A deep dive into optimizing MNIST digit predictions using semi-supervised learning with just 100 labeled samples. Utilizes pseudo-labels to bridge the gap between labeled and unlabeled data, leveraging TensorFlow for model implementation. A compact showcase of enhancing model accuracy with minimal labeled data.

Notifications You must be signed in to change notification settings

AbirOumghar/MNIST-SemiSupervised-PseudoLabels

Repository files navigation

MNIST-SemiSupervised-PseudoLabels

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.

Methodology

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.

Key Components

  • 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.

Implementation

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.

About

A deep dive into optimizing MNIST digit predictions using semi-supervised learning with just 100 labeled samples. Utilizes pseudo-labels to bridge the gap between labeled and unlabeled data, leveraging TensorFlow for model implementation. A compact showcase of enhancing model accuracy with minimal labeled data.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published