LatentWeights DoNotExist: RethinkingBinarized NeuralNetworkOptimization