<ul data-eligibleForWebStory="true"><li>Researchers have introduced a meta-learning approach to handle learning from multiple noisy annotators.</li><li>The method addresses scenarios like crowdsourcing where supervised learning labels come from different annotators with varied skills and biases.</li><li>Existing methods usually demand a large amount of noisy labeled data to train accurate classifiers, which may not always be available.</li><li>To mitigate data scarcity, the new approach leverages labeled data from related tasks.</li><li>It involves embedding examples into a latent space using a neural network and constructing a probabilistic model to learn task-specific classifiers while estimating annotators' abilities.</li><li>The neural network is meta-learned to optimize test classification performance with a small amount of labeled data by adapting the classifier using the expectation-maximization (EM) algorithm.</li><li>The EM algorithm's steps are computed efficiently and backpropagated through the neural network for meta-learning.</li><li>The effectiveness of the method is demonstrated using both synthetic noise and real-world crowdsourcing datasets.</li></ul>

Meta-learning Representations for Learning from Multiple Annotators

Discover more