Semi-supervised Vs Self-supervised Learning

Other Types of machine learning systems under the Training Supervision category

Photo by Arseny Togulev on Unsplash

Semi-supervised learning

Labeling data takes a lot of time and effort, so in many cases, we have a lot of data without labels and only a small amount with labels. Semi-supervised learning can work with both labeled and unlabeled data. For example, in Google Photos, when we upload pictures of our family and friends, the app automatically groups photos of the same person. If we label just a few of them with names, the app can figure out who’s who in the rest of the photos. This helps when we’re searching for pictures later.

Semi-supervised learning usually combines both unsupervised learning (like grouping similar faces) and supervised learning (like using the few names we added as labels). First, it finds similar groups, then it gives each group a label based on the few labeled examples. After labeling the whole dataset, it is possible to use any supervised learning algorithm.

Self-supervised learning

Self-supervised learning algorithms can generate a fully labeled dataset from a complete unlabeled one. Once the data is labeled, we can use any supervised learning algorithm. For example, suppose we have a big set of images. We can hide a small part of each image and train a model to guess what the hidden part looks like. The hidden image is the input, and the full original image is the label. This kind of model can be useful for fixing damaged photos or removing unwanted objects.

But usually, we don’t use the self-supervised model for the final task. Instead, we use it as a starting point and then tweak or fine-tune it for a specific job. For example, the goal is to build a model that can recognize different types of animals. So, we take the self-trained model and fine-tune it using a smaller labeled dataset of animal photos (e.g., dogs, cats, birds). Now the model performs well on this specific task, even though it started with no labels at all.