The connection and difference between self-supervised learning, supervised learning and unsupervised learning (classification of self-supervised learning framework)

Self-supervised learning (SSL) is a form of supervised learning that does not require manual input of data labels. It obtains results through models that independently analyze data, label and classify information without any human intervention. This method can reduce the workload of manual annotation, improve training efficiency, and perform well on large-scale data sets. SSL is a promising learning method that can be applied in various fields such as computer vision and natural language processing.

Self-supervised learning is an unsupervised learning method that uses unlabeled data to generate supervisory signals. Simply put, it trains the model by generating high-confidence data labels and then uses these labels in the next iteration. In each iteration, the ground truth based on the data labels changes. This approach can effectively utilize unlabeled data to improve model performance.

The relationship between self-supervised learning, supervised learning and unsupervised learning

Supervised learning requires the use of high-quality manually labeled data to train the model, and Adjust model weights.

Self-supervised learning uses data and automatically generated labels to train the model, with no real labels at the beginning.

Unsupervised learning works on data sets for which no labels are available. This learning paradigm attempts to understand the data provided without using labels at any stage of its training.

Self-supervised learning is a subset of unsupervised learning, both of which only provide unstructured data. But unsupervised learning works on clustering, grouping, and dimensionality reduction, whereas self-supervised learning performs tasks like classification, segmentation, and regression like other supervised models.

Framework type of self-supervised learning

Energy-based model (EBM)

Based on Models of energy attempt to use mathematical functions to calculate compatibility between two given inputs. When given two inputs, if the EBM produces a low energy output, it means the inputs have high compatibility; a high energy output means high incompatibility.

Joint Embedding Architecture

The joint embedding architecture is a two-branch network where the structure of each branch is the same. Each branch is given two inputs to compute their individual embedding vectors. There is a module at the head of the network that takes as input two embedding vectors and calculates their distance in the latent space.

So when two inputs are similar to each other, the calculated distance should be small. Network parameters can be easily tuned to ensure that inputs in the latent space are close to each other.

Contrastive learning

In contrastive learning SSL, inputs called "anchors" such as text, images, and videos are The clips are compared with positive and negative examples to train the model. Positive samples refer to samples that belong to the same distribution as the anchor point, while negative samples refer to samples that belong to a different distribution than the anchor point.

Non-contrast self-supervised learning (NC-SSL)

Non-contrast self-supervised learning (NC-SSL) is a learning paradigm. Only positive sample pairs are used to train the model, which is different from contrastive learning that uses both positive and negative sample pairs. However, NC-SSL has been shown to be able to learn non-trivial representations with only positive pairs using additional predictors and stopping gradient operations.

The above is the detailed content of The connection and difference between self-supervised learning, supervised learning and unsupervised learning (classification of self-supervised learning framework). For more information, please follow other related articles on the PHP Chinese website!