Skip to main content

Semi Supervised Learning

Created on April 25|Last edited on May 1

Useful links:

Steps for pseudo label method:

  1. Build a classifier on the labeled data that you have in the dataset.
  2. Use this classifier to predict the labels of the unlabelled data present in the dataset. You also check the **confidence level** of the prediction made.
  3. Add those observations made on the prediction to the training data to be a part of labelled data. These are called as pseudo labelled data.
  4. While choosing the observation, we have to carefully pick by analyzing the level of confidence(a design parameter). If you choose high level of confidence data as your pseudo labelled data, the data will not add much value to the training set because adding similar information to the dataset will not give much knowledge and classifier will be weak. On the other hand, choosing a low level of confidence pseudo labelled data will give us wrong observations leading to a bad dataset.
  5. Once you have chosen your observation and added the pseudo labelled data to the labelled data, use this augmented dataset to further train your model and use this model.
As we are using the unsupervised data to augment the training data for supervised learning, this comes in between supervised and unsupervised, hence the name semi-supervised learning.

Methods for Semi Supervised Learning in Computer Vision:

Self Training:

  1. Pseudo-label
  2. Noisy Student

Consistency Regularization:

  1. pi-model
  2. Temporal Ensembling
  3. Mean Teacher
  4. Virtual Adversarial Training
  5. Unsupervised Data Augmentation

Hybrid Methods:

  1. MixMatch
  2. FixMatch


Ayush Thakur
Ayush Thakur •  
Consistency Regularization What's Consistency Regulaization? Do you have any link to read?
1 reply
Ayush Thakur
Ayush Thakur •  
Nice.
Reply