Weakly supervised label learning flows.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Supervised learning demands extensive labeled data, which is often expensive to acquire.
• Weakly supervised learning offers an alternative by utilizing cheaper, approximate labels (weak signals).
• Existing methods often learn deterministic functions, which may not fully capture label uncertainty.

Purpose of the Study:

• To introduce a general framework for weakly supervised learning problems called label learning flows (LLF).
• To develop a novel generative model based on normalizing flows for label learning.
• To optimize conditional likelihoods of data labelings within weak signal constraints.

Main Methods:

• Developed label learning flows (LLF), a generative model utilizing normalizing flows.
• Proposed a training method that optimizes conditional likelihoods inversely, avoiding direct label estimation.
• Employed a sampling algorithm for making predictions after model training.

Main Results:

• Applied LLF to three distinct weakly supervised learning tasks.
• Demonstrated that LLF outperforms several established baseline methods.
• Validated the effectiveness of the generative approach in weakly supervised settings.

Conclusions:

• Label learning flows (LLF) provide a powerful and flexible framework for weakly supervised learning.
• The generative approach effectively handles label uncertainty and optimizes label likelihoods.
• LLF offers a promising alternative to traditional deterministic methods in scenarios with limited labeled data.