On Symmetry and Initialization for Neural Networks
This work addresses the challenge of improving training efficiency and generalization for neural networks in symmetric function learning, representing an incremental theoretical advance.
The paper tackles the problem of efficiently training neural networks to learn symmetric functions by proposing specific initial conditions that enable standard SGD to achieve generalization guarantees, which is empirically shown to fail with random initialization.
This work provides an additional step in the theoretical understanding of neural networks. We consider neural networks with one hidden layer and show that when learning symmetric functions, one can choose initial conditions so that standard SGD training efficiently produces generalization guarantees. We empirically verify this and show that this does not hold when the initial conditions are chosen at random. The proof of convergence investigates the interaction between the two layers of the network. Our results highlight the importance of using symmetry in the design of neural networks.