Unsupervised Learning

• Unsupervised learning
  • Training LLMs like GPT-4
    • considered as UL, or commonly as self-supervised learning (subset of UL)
    • dont have explicit labels or target outputs, but rather learning to predict next word in a sentence, given the previous words
    • using the input text itself as both input and output
  • data: given input x, but not output labels y
    • eg. data on patients, tumor’s size, patient’s age
  • goal: find sth interesting / patterns / structure in unlabeled data, not whether benign or malignant
  • clustering algo:
    • visualization
    • place data in different clusters
    • eg. gg news, group related stories together by keywords, in which the algo figure out by itself without supervision
  • will also learn
    • anomaly detection: find unusual data points
    • dimensionality reduction: compress data using fewer numbers