[Paper Review] Deep Batch Active Learning by Diverse, Uncertain Gradient Lower Bounds

Jordan T. Ash, Chicheng Zhang+3 more|arXiv (Cornell University)|Jun 9, 2019

Machine Learning and AlgorithmsComputer Science46 references280 citations

TL;DR

BADGE selects batches by combining gradient-based uncertainty with diversity, using k-means++ over gradient embeddings to query labels without extra hyperparameters, showing robust performance across architectures and batch sizes.

ABSTRACT

We design a new algorithm for batch active learning with deep neural network models. Our algorithm, Batch Active learning by Diverse Gradient Embeddings (BADGE), samples groups of points that are disparate and high-magnitude when represented in a hallucinated gradient space, a strategy designed to incorporate both predictive uncertainty and sample diversity into every selected batch. Crucially, BADGE trades off between diversity and uncertainty without requiring any hand-tuned hyperparameters. We show that while other approaches sometimes succeed for particular batch sizes or architectures, BADGE consistently performs as well or better, making it a versatile option for practical active learning problems.

Motivation & Objective

• Motivate label-efficient learning for deep neural networks in practical settings.
• Develop a batch active learning algorithm that works across architectures, batch sizes, and datasets without hand-tuned hyperparameters.
• Encode both uncertainty and diversity into batch selection via gradient-based representations.
• Provide a scalable sampling method that avoids costly probabilistic sampling while preserving batch quality.

Proposed method

• Compute a gradient embedding for each unlabeled example by taking the gradient of the cross-entropy loss with respect to the final layer, using the model’s current predicted label as the surrogate true label.
• Use the length of the gradient embedding as a proxy for uncertainty and potential update magnitude.
• Select batches via k-means++ seeding on the set of gradient embeddings to promote both high magnitude and diversity among batch points.
• Iteratively query labels for selected batch points, retrain the model, and repeat for T iterations.
• Prove that the gradient embedding norm lower-bounds the true gradient norm for the final layer under common network structures.
• Compare BADGE to several baselines across multiple architectures (MLP, ResNet, VGG) and datasets (SVHN, CIFAR-10, MNIST, OpenML datasets).

Experimental results

Research questions

• RQ1 Does BADGE consistently outperform or match state-of-the-art batch active learning methods across different architectures, batch sizes, and datasets?
• RQ2 Can a hyperparameter-free, gradient-embedding based approach capture both uncertainty and diversity effectively in batches?
• RQ3 Is k-means++ sampling of gradient embeddings a computationally efficient and robust surrogate for more complex samplers like k-DPP?
• RQ4 How does BADGE perform relative to representative sampling (diversity) and uncertainty sampling baselines in practical deep learning settings?

Key findings

• BADGE consistently performs as well as or better than the best baselines across architecture, batch size, and dataset settings.
• Gradient embeddings provide a conservative measure of uncertainty and a direction for potential updates, enabling effective batch selection.
• k-means++ on gradient embeddings yields diverse, high-magnitude batches comparable to or better than k-DPP sampling, with improved run-time.
• Diversity-based methods alone can fail on complex data or with limited architectural priors, while BADGE maintains robustness by combining uncertainty and diversity.
• In experiments, BADGE often matches or exceeds performance of coreset, margin, entropy, and random baselines, particularly at smaller batch sizes and with convolutional architectures.