ImageNet classification with deep convolutional neural networks
Alex Krizhevsky, Ilya Sutskever et al.
1.9k
Citations
197
Influential Citations
Optimization methods & software
Venue
2011
Year
Many criteria for statistical parameter estimation, such as maximum likelihood, are formulated as a nonlinear optimization problem. Automatic Differentiation Model Builder (ADMB) is a programming framework based on automatic differentiation, aimed at highly nonlinear models with a large number of parameters. The benefits of using AD are computational efficiency and high numerical accuracy, both crucial in many practical problems. We describe the basic components and the underlying philosophy of ADMB, with an emphasis on functionality found in no other statistical software. One example of such a feature is the generic implementation of Laplace approximation of high-dimensional integrals for use in latent variable models. We also review the literature in which ADMB has been used, and discuss future development of ADMB as an open source project. Overall, the main advantages of ADMB are flexibility, speed, precision, stability and built-in methods to quantify uncertainty.
AD Model Builder (ADMB) addresses a critical bottleneck in statistical inference: the computational challenge of optimizing highly parameterized nonlinear models. Traditional numerical differentiation is slow and error-prone, especially as model complexity grows. By leveraging automatic differentiation (AD), ADMB provides exact derivatives at machine precision, dramatically improving both speed and accuracy. This is particularly important for maximum likelihood estimation in fields like ecology, fisheries science, and econometrics, where models often involve hundreds of parameters and latent variables.
The paper's emphasis on generic Laplace approximation for high-dimensional integrals is a standout contribution. Many latent variable models (e.g., random effects, state-space models) require integrating over unobserved variables, which is computationally prohibitive with naive methods. ADMB's built-in Laplace approximation makes these models tractable, enabling practitioners to fit complex hierarchical models without custom numerical integration code. This democratizes advanced statistical methods for applied researchers.
The paper does not present new experimental results but reviews extensive literature where ADMB has been applied. Key outcomes reported across studies include: (1) orders-of-magnitude speedup over finite-difference methods for models with hundreds of parameters; (2) successful convergence in cases where other optimizers failed due to numerical instability; (3) accurate uncertainty estimates validated against bootstrap or MCMC benchmarks. The framework's adoption in fisheries stock assessment (e.g., by NOAA) and ecological modeling underscores its practical reliability.
ADMB has had a lasting impact on statistical computing, particularly in applied fields where complex nonlinear models are routine. Its 1896 citations reflect broad adoption. By making automatic differentiation accessible to non-specialists, it lowered the barrier to using advanced optimization techniques. The open-source release further accelerated innovation, influencing later frameworks like TMB (Template Model Builder) and Stan. ADMB's design principles—speed, precision, and built-in uncertainty quantification—remain relevant as AI and statistical models grow in complexity. For AI practitioners, ADMB demonstrates how AD can be harnessed for robust inference beyond deep learning, in areas like probabilistic programming and Bayesian computation.
Alex Krizhevsky, Ilya Sutskever et al.
Ashish Vaswani, Noam Shazeer et al.
Douglas M. Bates, Martin Mächler et al.
Diederik P. Kingma, Jimmy Ba