The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Learning with Limited Labeled Data with Shioulin Sam - TWiML Talk #255
Apr 22, 2019
In this conversation, Shioulin Sam, a Research Engineer at Cloudera Fast Forward Labs, sheds light on the challenges of machine learning with limited labeled data. She discusses active learning as a powerful approach to enhance model accuracy while reducing costly manual labeling. The discussion delves into measuring uncertainty in deep neural networks and effective strategies for implementing active learning across enterprises. Shioulin also highlights the significance of integrating labeled data into models to drive optimal performance in real-world applications.
AI Snips
Chapters
Transcript
Episode notes
Active Learning Strategies
- Active learning strategies identify difficult data points for labeling.
- Classical strategies often use uncertainty, like distance to decision boundary, while newer ones adapt for deep learning's complexity.
Uncertainty in Deep Learning
- Prediction probability as uncertainty measure works in classical active learning but not for deep neural networks.
- Deep networks can misclassify with high confidence due to their complexity, as shown by adversarial examples.
Dropout for Uncertainty
- Dropout, typically a regularization technique, can estimate posterior weight distribution in Bayesian neural networks.
- Leaving dropout on during inference provides weight samples, simplifying uncertainty calculation in active learning.