Pruning Neural Networks as a Game: An Equilibrium Approach

This paper introduces a novel perspective on neural network pruning, framing it as a game-theoretic problem. Instead of relying on heuristics, it models network components as players in a non-cooperative game, where sparsity emerges as an equilibrium outcome. This approach offers a principled explanation for pruning behavior and leads to a new pruning algorithm. The focus is on establishing a theoretical foundation and empirical validation of the equilibrium phenomenon, rather than extensive architectural or large-scale benchmarking.

• •Proposes a game-theoretic framework for neural network pruning.
• •Sparsity emerges as an equilibrium outcome.
• •Offers a principled explanation for pruning.
• •Develops a new equilibrium-driven pruning algorithm.
• •Achieves competitive sparsity-accuracy trade-offs.