Trouble using input parameters on a custom deep learning layer

I'm trying to perform feature selection methods on a deep learning layer. My initial idea was to use a learnable parameter that I inputted in the custom layer, with all the indexes of the vector I want to select with a '1' value. But when I call the layer.Variable, all the indexes changed its value and are consistent with what I initially inputted.

classdef globalindexPooling < nnet.layer.Layer
  properties (Learnable)
      Alpha
  end

    methods
        function layer = globalindexPooling(numChannels, name) 
            % Initialize scaling coefficient
            layer.Alpha = numChannels;
        end
        function Z = predict(layer, X)
            input_size = [size(X,1) size(X,2) size(X,3)];
            N = size(X,4);
            X_reshaped = reshape(X,[prod(input_size) N]);  
            Z = X_reshaped(logical(gather(layer.Alpha)),:);
            Z = reshape(Z,[6 6 256 N]);
        end
        function [Z, memory] = forward(layer, X) 
            input_size = [size(X,1) size(X,2) size(X,3)];
            N = size(X,4);
            X_reshaped = reshape(X,[prod(input_size) N]);  
            Z = X_reshaped(logical(gather(layer.Alpha)),:);
            Z = reshape(Z,[6 6 256 N]); 
            memory = [];
        end
        function [dLdX,dLdalpha] = backward(layer, X, Z, dLdZ,~) 
            input_size = [size(X,1) size(X,2) size(X,3)];
            N = size(X,4);
            dLdZ = reshape(dLdZ,[6*6*256 N]);
            dLdX_prima = zeros([prod(input_size) N],'like',X);
            dLdX_prima(logical(gather(layer.Alpha)),:) = dLdZ;
            dLdX = reshape(dLdX_prima,[input_size N]);
            dLdalpha = ones(size(gather(layer.Alpha)),'like',X); 
        end

The input is a logical vector with ones in the selected indexes, but I got all 0.0784 and -0.9216, sometimes the first values being the original ones and other not. I tried using round() but since the assignment to each of them is inconsistent, it didn't work.