Improving the performance of fermionic neural networks with the Slater exponential Ansatz