Unsupervised learning

Unsupervised learning is a type of machine learning where algorithms find patterns in data without any labels. This is different from supervised learning, where data comes with labels to guide the learning process. Other types of learning, like weak or semi-supervised learning, use some labeled data, while self-supervised learning is sometimes considered a form of unsupervised learning.

In unsupervised learning, the data is often collected from sources like the internet with little processing, making it cheaper than supervised learning, where labeled datasets are created manually.

There are special algorithms for unsupervised learning, such as k-means for grouping similar data, PCA for reducing data complexity, and autoencoders for learning useful representations. With the rise of deep learning, large-scale unsupervised learning is mostly done using neural networks trained with gradient descent and special training techniques.

Sometimes, models trained with unsupervised learning are used directly, but often they are adjusted for other tasks. For example, a model trained to generate text can later be fine-tuned for tasks like text classification. Similarly, autoencoders can learn useful features that help improve other models, such as those used in image generation.^[1][2]

1. ↑ Wu, Wu (2022). "Unsupervised Learnin" (PDF). Mathematisches Institut Universität Tübingen.
2. ↑ Liu, Xiao; Zhang, Fanjin; Hou, Zhenyu; Mian, Li; Wang, Zhaoyu; Zhang, Jing; Tang, Jie (2023-01). "Self-Supervised Learning: Generative or Contrastive". IEEE Transactions on Knowledge and Data Engineering. 35 (1): 857–876. doi:10.1109/TKDE.2021.3090866. ISSN 1558-2191. {{cite journal}}: Check date values in: |date= (help)