There are Moczulski et al. with their ACDC paper [1]. They describe a structured layer that can approximate any linear transform and discuss an optical implementation of it (see section 1.1 "lightning fast deep sell" :).
Our deep SELL offers several possibilities for analog physical implementation. Given
the great demand for fast low energy neural networks, the possibility of harnessing physical phenomena to perform efficient computation in deep networks is worthy of consideration.... (see paper)
Our deep SELL offers several possibilities for analog physical implementation. Given the great demand for fast low energy neural networks, the possibility of harnessing physical phenomena to perform efficient computation in deep networks is worthy of consideration.... (see paper)
[1] https://arxiv.org/pdf/1511.05946.pdf