Data Programming: Creating Large Training Sets, Quickly.

Area of Science:

• Machine Learning
• Artificial Intelligence

Background:

• Large labeled training sets are crucial for supervised learning and deep learning.
• Creating these labeled sets is often the most time-consuming and expensive aspect of machine learning applications.

Purpose of the Study:

• To introduce a new paradigm called data programming for programmatic training set creation.
• To address the challenges of noisy and conflicting labels generated by weak supervision strategies.

Main Methods:

• Users define labeling functions (programs) that provide weak supervision or domain heuristics.
• A generative model represents the labeling process, allowing for denoising of the training set.
• Discriminative loss functions are modified to be noise-aware.

Main Results:

• The proposed method can theoretically recover generative model parameters in certain settings.
• Experimental results on the 2014 TAC-KBP Slot Filling challenge show improved performance.
• Applying data programming to an LSTM model yielded a significant F1 score improvement over a baseline.

Conclusions:

• Data programming offers a viable solution for efficient and effective training data generation.
• The method enhances the performance of various discriminative models, including LSTMs.
• Initial user studies suggest data programming simplifies machine learning model creation for non-experts, especially with limited data.