[Paper Review] Improved Algorithms for Neural Active Learning

Yikun Ban, Yuheng Zhang+3 more|arXiv (Cornell University)|Oct 2, 2022

Machine Learning and Algorithms4 citations

TL;DR

This paper proposes I-NeurAL, a novel neural active learning algorithm that improves regret bounds by introducing new population loss-based regret metrics, leverages full feedback and warm-start mini-batch SGD for efficiency, and achieves instance-dependent regret with no curse of dimensionality. It matches Bayes-optimal performance under hard-margin conditions and outperforms SOTA baselines empirically across multiple datasets.

ABSTRACT

We improve the theoretical and empirical performance of neural-network(NN)-based active learning algorithms for the non-parametric streaming setting. In particular, we introduce two regret metrics by minimizing the population loss that are more suitable in active learning than the one used in state-of-the-art (SOTA) related work. Then, the proposed algorithm leverages the powerful representation of NNs for both exploitation and exploration, has the query decision-maker tailored for $k$-class classification problems with the performance guarantee, utilizes the full feedback, and updates parameters in a more practical and efficient manner. These careful designs lead to an instance-dependent regret upper bound, roughly improving by a multiplicative factor $O(\log T)$ and removing the curse of input dimensionality. Furthermore, we show that the algorithm can achieve the same performance as the Bayes-optimal classifier in the long run under the hard-margin setting in classification problems. In the end, we use extensive experiments to evaluate the proposed algorithm and SOTA baselines, to show the improved empirical performance.

Motivation & Objective

• To address the lack of theoretical performance guarantees in neural active learning, especially in non-parametric streaming settings.
• To overcome limitations in prior work, such as reliance on pseudo-regret metrics and inefficient training with random re-initialization.
• To eliminate the curse of input dimensionality $d$ and effective dimensionality $\widetilde{d}$ in regret bounds.
• To design a practical, efficient algorithm using warm-start mini-batch SGD and full feedback for improved generalization.
• To achieve performance matching the Bayes-optimal classifier under hard-margin assumptions.

Proposed method

• Introduces two new regret metrics based on minimizing population loss, directly reflecting generalization capability.
• Designs a neural exploration strategy with a query decision mechanism tailored for $k$-class classification, with theoretical performance guarantees.
• Employs full feedback by updating model parameters using all available labels, enhancing learning efficiency.
• Uses warm-start mini-batch SGD, initializing from the previous round's parameters to improve training efficiency and convergence.
• Provides an instance-dependent regret upper bound that improves by a multiplicative $O(\log T)$ factor and removes dependence on $d$ and $\widetilde{d}$.
• Theoretically shows that under hard-margin conditions, the model achieves Bayes-optimal performance after $\mathcal{O}(\log T)$ queries.

Experimental results

Research questions

• RQ1 Can we design regret metrics in neural active learning that directly reflect generalization performance rather than relying on pseudo-regret?
• RQ2 How can we improve training efficiency in neural active learning while maintaining theoretical guarantees?
• RQ3 Can we eliminate the curse of dimensionality in regret bounds for over-parameterized neural networks?
• RQ4 Does the proposed algorithm achieve performance comparable to the Bayes-optimal classifier under realistic data assumptions?
• RQ5 How does the algorithm perform empirically under varying label budgets and real-world data distributions?

Key findings

• I-NeurAL achieves a regret bound that improves by a multiplicative factor of $O(\log T)$ compared to prior work, with no dependence on input dimensionality $d$ or effective dimensionality $\widetilde{d}$.
• On the Phishing dataset with 3% label budget, I-NeurAL achieves 94.22% test accuracy, outperforming the next best method (93.46% for Margin) and significantly surpassing NeuAL-NTK-F (54.69%).
• With 10% label budget, I-NeurAL attains 95.64% accuracy on Phishing, 97.90% on IJCNN, and 99.30% on Fashion MNIST, exceeding all baselines.
• On CIFAR-10 with 50% label budget, I-NeurAL reaches 92.30% accuracy, outperforming ALPS (90.85%) and NeuAL-NTK-D (90.55%).
• Sensitivity analysis shows that $\gamma = 6$ or $7$ yields optimal performance, balancing query efficiency and model confidence.
• Ablation studies confirm that I-NeurAL maintains superior performance across all label budget levels (3%, 10%, 20%, 50%), especially under low-budget conditions.