Useful, but let's don't work ourselves into a frenzy
If this is what I think it is, an optical version of a neural net built from a block of glass fibres, it has to be hard-wired.
That means that it has to be hard-wired for a specific purpose. At low levels this isn't all that bad: we have hardwired edge and area detection for our own eyes, providing additional information derived from a field of pixels which is useful for recognizing what we see.
Scaling that up won't be one of those square-of-the-number-of-pixels things since that sort of recognition depends on nearby pixels rather than correlating information from distant parts of the visual field.
So far, so good.
The problem comes with hard-wiring the higher levels of recognition.
You want to recognize numbers and letters? Roman letters? Cyrillic? Arabic? Telugu?
You want to recognize faces? Whose faces? Maybe worth hard-wiring to recognize specific features, but beyond that not so much.
So at some point someone has either to create a *programmable* optical neural net to sit atop this device, or to use the device to enhance the input to more conventional software.
Sounds useful, but this isn't the be-all and end-all of image recognition. Recognize it for what it is, to wit a useful step forward, but let's don't get all dizzy over it.