Multiple-Population Moment Estimation: Exploiting Interpopulation Correlation for Efficient Moment Estimation in Analog/Mixed-Signal Validation

Chenjie Gu, Manzil Zaheer, Xin Li

TL;DR

This paper introduces Multiple-Population Moment Estimation (MPME), a novel Bayesian approach that significantly boosts the accuracy of estimating circuit performance moments (like mean and variance) when dealing with very limited data. By cleverly exploiting correlations between data from different circuit configurations or "corners," MPME achieves up to a 2x reduction in estimation error, leading to potentially substantial savings in validation time and cost for complex analog/mixed-signal circuits.

What Problem Does This Paper Address?

Validating complex analog and mixed-signal circuits is a critical but challenging step in the design process. This involves estimating key performance metrics, such as the probability of failure or parametric yield, which are derived from the statistical moments (mean and variance) of circuit performance. These metrics guide design optimization and product qualification. The problem is particularly acute because obtaining measurements or running simulations for these circuits is extremely time-consuming and expensive. This often forces designers to work with very small sample sizes (sometimes as few as 1 to 10 data points per configuration). Traditional methods struggle with such small sample sizes, leading to biased and unreliable estimates. Furthermore, validation must often be performed across numerous "corners" (e.g., different process, voltage, temperature settings) and configurations (e.g., different board layouts, input patterns), making it impractical to gather enough data for each individual scenario.

Methods

The core innovation is the Multiple-Population Moment Estimation (MPME) method. MPME tackles the small sample size problem by treating data from different populations (defined by specific circuit configurations or corners) not as independent, but as correlated. This is achieved through a hierarchical Bayesian framework.

Here's how it works:

  1. Generative Graphical Model: MPME models the data using a generative graphical model. This model specifies a joint prior distribution over the moments (mean and variance) of different populations. This implicitly captures the underlying correlations between these populations.
  2. Two-Step Estimation:
    • Prior Learning (MLE): The first step involves learning the parameters of this joint prior distribution from the available data across all populations using Maximum Likelihood Estimation (MLE). This step essentially learns how the moments of different populations are related.
  3. MAP Estimation: In the second step, Maximum A Posteriori (MAP) estimation is used to obtain the moment estimates for each individual population. This leverages the learned prior distribution, effectively "borrowing strength" from other populations to improve the estimate for the current one.

The paper proposes two specific prior distributions: the Uniform Prior (UNI) and the Normal-Inverse-Chi-Squared Prior (NIX). The NIX prior is highlighted for its computational efficiency due to its conjugate nature, allowing for closed-form MAP solutions.

Datasets

The paper evaluates MPME on:

  • Synthetic Datasets: Two synthetic examples are generated to systematically study the impact of varying the number of populations (P) and sample sizes (N) on estimation accuracy.
  • Industrial Dataset: Post-silicon measurement data from a commercial high-speed I/O link is used to demonstrate the practical effectiveness of MPME.

The paper does not introduce new benchmark datasets but uses existing measurement data.

Key Insights and Ideas

  • Exploiting Inter-Population Correlation is Key: The fundamental insight is that data from different circuit configurations/corners are not independent but share underlying correlations. MPME's success hinges on effectively leveraging this correlation.
  • "Borrowing Strength" for Small Samples: By modeling a joint prior, MPME allows information from populations with more data (or better data) to inform estimates for populations with very little data. This is analogous to how James-Stein estimators improve upon sample means by shrinking estimates towards a common mean.
  • Bayesian Framework as a Natural Fit: The hierarchical Bayesian approach provides a principled way to model these inter-population dependencies and perform robust inference, especially under data scarcity.
  • NIX Prior Offers Efficiency and Accuracy: The Normal-Inverse-Chi-Squared prior is particularly advantageous because it's a conjugate prior, leading to computationally efficient MAP estimation and often yielding better overall accuracy compared to the Uniform prior, though the Uniform prior can be better for extreme corner cases.

Key Results

  • Significant Accuracy Improvement: MPME consistently outperforms traditional sample mean and variance estimators, achieving an average error reduction of up to 2x across various scenarios.
  • Effectiveness with Small Sample Sizes: The benefits of MPME are most pronounced when sample sizes are extremely small (e.g., N=5).
  • Scalability with Populations: As the number of populations (P) increases, MPME's error decreases roughly as 1/√P, demonstrating its scalability.
  • Practical Validation Impact: For a commercial high-speed I/O link, MPME achieved similar accuracy with about 50% fewer samples compared to traditional methods, directly translating to reduced validation time and cost.

Limitations and Future Work

  • Potential for Bias in Extreme Cases: The authors acknowledge that in theoretically pathological scenarios with many populations and extreme outliers, MPME might introduce bias by shrinking estimates towards a common mean. However, they argue these cases are unlikely in practice and outliers can be handled with pre-processing.
  • Convexity of Optimization: The optimization problems in MPME might not always be convex, meaning numerical optimizers might find local optima rather than the global ones. However, good initial guesses from the data help mitigate this.
  • Non-Gaussian Distributions and Higher-Order Moments: While the paper focuses on Gaussian distributions and estimating the first two moments, the authors suggest that MPME can be extended to non-Gaussian distributions (especially exponential family) and potentially higher-order moments, though this would require more complex modeling and might face computational challenges.
  • Empirical Priors: The paper briefly mentions using empirical priors based on designer knowledge but cautions against their use due to potential bias, suggesting cross-validation as a check.

Summary updated

Abstract

Moment estimation is an important problem during circuit validation, in both presilicon and postsilicon stages. From the estimated moments, the probability of failure and parametric yield can be estimated at each circuit configuration and corner, and these metrics are used for design optimization and making product qualification decisions. The problem is especially difficult if only a very small sample size is allowed for measurement or simulation, as is the case for complex analog/mixed-signal circuits. In this paper, we propose an efficient moment estimation method, called multiple-population moment estimation (MPME), that significantly improves estimation accuracy under small sample size. The key idea is to leverage the data collected under different corners/configurations to improve the accuracy of moment estimation at each individual corner/configuration. Mathematically, we employ the hierarchical Bayesian framework to exploit the underlying correlation in the data. We apply the proposed method to several datasets including postsilicon measurements of a commercial high-speed I/O link, and demonstrate an average error reduction of up to 2×, which can be equivalently translated to significant reduction of validation time and cost.
BibTeX
@article{gu2014multiple,
  author = {Chenjie Gu and Manzil Zaheer and Xin Li},
  title = {{Multiple-Population Moment Estimation: Exploiting Interpopulation Correlation for Efficient Moment Estimation in Analog/Mixed-Signal Validation}},
  journal = {IEEE Trans. Comput. Aided Des. Integr. Circuits Syst.},
  year = {2014},
  url = {https://openreview.net/forum?id=MsCE6I4gJ2},
}
Figure