The Lottery Ticket Hypothesis (LTH) asserts that a randomly initialised overparameterised Deep Neural Network (DNN) contains a sparse subnetwork that, when trained (up to) the same amount as the original network, performs just as well. So-called winning tickets are vastly more efficient to train than dense networks; however, finding such tickets currently relies on pre-training an overparameterised network via Iterative Magnitude Pruning (IMP). Due to the increasing demand for computing resources, there are major incentives to find good search procedures for winning tickets (which can drastically reduce the training time required to solve a range of problems). In this paper, we propose a new method for the evolution of sparse DNN initialisations that generates results commensurate with winning ticket search procedures: we refer to the method as Neuroevolution Ticket Search (NeTS). By training an overparameterised network using gradient descent for use as a baseline only, we show that NeTS quickly converges to near state-of-the-art performance in terms of network size and trainability. Additionally, we find that NeTS appears to be a competitive alternative to IMP in terms of wall clock time.
|Title of host publication||International Conference on Learning Representations 2023|
|Subtitle of host publication||Workshop on Sparsity in Neural Networks|
|Number of pages||11|
|Publication status||Accepted/In press - 3 Mar 2023|
- machine learning
- Evolutionary algorithms