You could just ensemble a number of neural networks by averaging across dimensions.
That is to miss some opportunities for neurons to help each other out, so to say.
Or you can take a weak learner view point and then find a better way to combine those into stronger learners. Or you take a quantization viewpoint and find ways to get finer granulation.
Anyway better to use Random Projections to ensemble:
https://sciencelimelight.blogspot.com/2025/08/sum-then-project-vs-project-then-sum.html