Stochastic Approximate Gradient Descent via the Langevin Algorithm
This addresses a bottleneck in machine learning for problems with intractable gradients, offering a practical alternative to general-purpose sampling methods, though it appears incremental as it builds on existing Langevin techniques.
The paper tackles the problem of optimizing objectives where unbiased stochastic gradients are not easily available, by introducing the stochastic approximate gradient descent (SAGD) algorithm, which uses the Langevin algorithm to construct biased but asymptotically accurate gradients and shows experimental performance in tasks like expectation-maximization and variational autoencoders.
We introduce a novel and efficient algorithm called the stochastic approximate gradient descent (SAGD), as an alternative to the stochastic gradient descent for cases where unbiased stochastic gradients cannot be trivially obtained. Traditional methods for such problems rely on general-purpose sampling techniques such as Markov chain Monte Carlo, which typically requires manual intervention for tuning parameters and does not work efficiently in practice. Instead, SAGD makes use of the Langevin algorithm to construct stochastic gradients that are biased in finite steps but accurate asymptotically, enabling us to theoretically establish the convergence guarantee for SAGD. Inspired by our theoretical analysis, we also provide useful guidelines for its practical implementation. Finally, we show that SAGD performs well experimentally in popular statistical and machine learning problems such as the expectation-maximization algorithm and the variational autoencoders.