Problem defining a Neural Network with gradient as Loss function

Specific Domains Machine Learning
1

Hi, I’m trying to setup a simple network who’s Loss Function is the Laplacian of the NN with respect to the gradient. Then, the loss function is:

function 𝕃1(x)
    f(j) =  NNsolver(j) #Output of the NN
    ∇²(n) = Zygote.diaghessian(n->f(n)[1],n) #Diagonal hessian (since 1D)
    ∇²(x)[1][1] #Output a scalar
end

And the training process is also pretty straight forward:

function trainer()
    for point=Rx
        p=[point]
        gs = gradient(params(NNsolver)) do # NN model
          𝕃1(p)
        end
        update!(opt, params(NNsolver), gs)
      end

end

Nonetheless, the function call outputs the following error (which I dont understand since I am not reshaping any array for example):

Mutating arrays is not supported -- called setindex!(::Vector{Float64}, _...)

Would really appreciate any help and thanks in advance!

2

Without an MWE that includes enough of NNsolver to trigger the error, we won’t have enough information to help you troubleshoot this. A full stack trace would be very much appreciated as well.