LP-NAS: Neural Architecture Search via Linear Programming

LP-NAS uses a linear programming framework to automate neural network architecture design by applying validation-loss gradients and training-loss Hessians. This method calculates an architecture update direction that optimizes generalization while maintaining the optimality of existing model parameters.