Stephen Dang
Comparison of Regression-Adjusted ABC and Neural Methods for Simulation-Based Inference
Many scientific models have intractable likelihoods, making classical Bayesian updating impractical. Simulation-Based Inference (SBI) addresses this by using simulations to approximate posterior distributions. Recent benchmarks have popularised neural methods such as Neural Posterior Estimation (NPE), but they seldom compare against Approximate Bayesian Computation with regression adjustment (RA-ABC), a theoretically grounded enhancement that corrects ABC draws and can be paired with flexible non-linear regressors. This project fills that gap.
We will conduct a systematic, head-to-head evaluation of RA-ABC (linear and non-linear variants) versus NPE across established SBI benchmark tasks under matched simulation budgets. The study will address three research questions:
When does RA-ABC match or outperform neural methods in accuracy and efficiency?
How do design choices (summaries, bandwidths, regression class) affect RA-ABC performance?
Can post-processing (recalibration) improve uncertainty calibration relative to both RA-ABC and NPE?
Methods and outputs:
Implement RA-ABC and NPE using open SBI benchmarks; assess accuracy (e.g., classifier two-sample tests, posterior-predictive checks), efficiency (runtime and simulations used), and calibration (coverage of credible intervals). Perform ablations on summaries and regression models, and report practical guidance on method selection. Deliverables include a reproducible codebase, benchmark results tables/figures, and a short paper-style report.
Impact:
The project will provide clear, evidence-based recommendations on when a lightweight, regression-adjusted ABC pipeline—possibly with recalibration—can substitute for more complex neural SBI, informing practitioners in fields such as ecology, genetics, and neuroscience.
Stephen Dang
Queensland University of Technology
Stephen Dang is an undergraduate student in the Bachelor of Data Science (Dean’s Scholar) program at the Queensland University of Technology (QUT), majoring in Statistics and Machine Learning. He currently holds a GPA of 6.675/7.000, reflecting consistent high achievement across advanced mathematics, statistical modelling, and computational analytics units.
Stephen has developed a strong interest in simulation-based inference, Bayesian methods, and statistical learning, motivated by their applications to complex real-world problems in climate science and economics. He has previously undertaken research projects under QUT’s Vacation Research Experience Scheme (2024/2025) and Mid-Year Research Experience Scheme (2025), where he worked on topics involving coral reef health modelling and climate–economy interactions. These experiences strengthened his skills in R, Python, and Bayesian computation, and sparked his enthusiasm for research at the intersection of statistics and artificial intelligence.
Alongside his academic pursuits, Stephen is involved in applied data analytics projects in the education technology sector, where he designs and evaluates large-scale digital learning analytics systems. This professional experience complements his research interests, particularly in developing data-driven models that connect theoretical inference with real-world decision-making.
