Path integral approach to random neural networks