Optimized Parameter-Efficient Deep Learning Systems via Reversible Jump Simulated Annealing

We utilize the non-convex optimization method simulated annealing enriched with reversible jumps to enable a model selection capacity for deep learning models in a model size aware context. By using simulated annealing enriched with reversible jumps, we can yield a robust stochastic learning of the hidden posterior distribution of the structure, simultaneously constructing a more focused and certain estimate of the structure, all while making use of all the data. Being based upon Markov-chain learning methods, we constructed our priors to favor smaller and simpler architectures, allowing us to converge on the set of globally optimal models that are additionally parameter-efficient, seeking low parameter count deep models that retain good predictive accuracy. We demonstrate the capability on standard image recognition with CIFAR-10, as well as performing model selection on time-series tasks, realizing networks with competitive performance as compared to competing non-convex optimization methods such as genetic algorithms, random search, and Gaussian process based Bayesian optimization, while being less than half the size.