Stacking as Accelerated Gradient Descent
This provides a theoretical foundation for a widely-used heuristic in deep learning and boosting, which is incremental but clarifies its underlying mechanism.
The paper tackles the problem of explaining the efficiency of stacking in training deep residual networks by theoretically linking it to Nesterov's accelerated gradient descent, and proves that stacking provides accelerated training for certain deep linear residual networks through a new potential function analysis.
Stacking, a heuristic technique for training deep residual networks by progressively increasing the number of layers and initializing new layers by copying parameters from older layers, has proven quite successful in improving the efficiency of training deep neural networks. In this paper, we propose a theoretical explanation for the efficacy of stacking: viz., stacking implements a form of Nesterov's accelerated gradient descent. The theory also covers simpler models such as the additive ensembles constructed in boosting methods, and provides an explanation for a similar widely-used practical heuristic for initializing the new classifier in each round of boosting. We also prove that for certain deep linear residual networks, stacking does provide accelerated training, via a new potential function analysis of the Nesterov's accelerated gradient method which allows errors in updates. We conduct proof-of-concept experiments to validate our theory as well.