Prior-aware and Context-guided Group Sampling for Active Probabilistic Subsampling

2026-07-09 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Prior-aware and context-guided Group-based Active Deep Probabilistic Subsampling (PGA-DPS) is a novel method enhancing active subsampling for deep learning tasks. It integrates a deterministic, prior-informed sampling pattern from training data with group-based active sampling via top-k selection, improving upon Active Deep Probabilistic Subsampling (A-DPS) which relies on top-1 sampling and doesn't fully use dataset priors. PGA-DPS offers a theoretical analysis demonstrating smoother optimization due to a smaller effective Lipschitz constant with group sampling. Empirical evaluations across classification (MNIST, CIFAR-10), image reconstruction (fastMRI knee), and segmentation (AeroRIT) tasks show PGA-DPS consistently outperforms A-DPS, DPS, and other state-of-the-art sampling methods, particularly at low sampling ratios. The code is available on GitHub.

Key takeaway

For Machine Learning Engineers optimizing data acquisition in resource-constrained environments, PGA-DPS offers a superior subsampling strategy. By combining prior-aware deterministic sampling with context-guided group sampling, you can achieve more robust optimization and significantly improve performance in tasks like medical imaging or hyperspectral analysis, especially at low sampling ratios. Consider implementing PGA-DPS to reduce data processing overhead and accelerate acquisition time in your applications.

Key insights

PGA-DPS combines prior-informed and group-based active sampling for robust, efficient deep learning optimization.

Principles

• Group sampling (DPS-top-k) yields a smaller effective Lipschitz constant than sequential (DPS-top-1).
• Smaller Lipschitz constants lead to smoother loss landscapes and more stable optimization.
• Integrating deterministic prior sampling with instance-adaptive active sampling improves performance.

Method

PGA-DPS starts with a deterministic, prior-informed subsampling pattern, then iteratively acquires additional samples using context-guided group sampling (DPS-top-k) rather than top-1.

In practice

• Apply PGA-DPS for efficient data acquisition in MRI, CT, and hyperspectral imaging.
• Use group sampling (top-k) over sequential (top-1) for more stable optimization.
• Consider performance discrepancy between DPS and A-DPS to guide prior sampling ratio (Ps).

Topics

• Active Subsampling
• Deep Probabilistic Subsampling
• Group Sampling
• Lipschitz Continuity
• Image Reconstruction
• Hyperspectral Imaging
• Medical Imaging

Code references

• B9Kang/PGADPS

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Prior-aware and Context-guided Group Sampling for Active Probabilistic Subsampling — Enhancing active probabilistic subsampling with prior knowledge and group sampling improves optimization and performance.
• State-of-art minibatches via novel DPP kernels: discretization, wavelets, and rough objectives — Wavelet-based DPPs and a novel discretization pipeline significantly enhance minibatch and coreset sampling efficiency for diverse ML tasks.
• Decentralized Proximal Stochastic Gradient Langevin Dynamics — DE-PSGLD enables decentralized MCMC sampling from constrained log-concave distributions using proximal regularization.
• Efficient Online Proportional Sampling with Applications to Smoothed Online Learning — A novel data structure enables efficient online proportional sampling in high-dimensional, dynamically evolving systems, achieving tight bounds.
• Bayesian Magnetic Resonance Joint Image Reconstruction and Uncertainty Quantification using Sparsity Prior Models and Markov Chain Monte Carlo Sampling — Bayesian MCMC with sparsity priors enables superior MR image reconstruction and uncertainty quantification, outperforming optimization and deep learning methods.
• Rank-Constrained Deep Matrix Completion for Group Recommendation — Group RC-DMC unifies low-rank regularization, linear encoder-decoders, and attention-based group modeling for accurate group recommendations.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.