One of the things that sets humans apart from machines is our ability to process the context of a situation and make intelligent decisions based on internal analysis and learned experiences.
Recent years have seen the development of new “smart” and artificially “intelligent” machine systems. While these do have intelligence based on analysing data and predicting outcomes, many intelligent machine networks struggle to contextualize information and tend to just create a general output that may or may not have situational context.
Whether we want to build machines that can make informed contextual decisions like humans can is an ethical debate for another day, but it turns out that neural networks can be equipped with recurrent feedback that allows them to process current inputs based on information from previous inputs. These so-called recurrent neural networks (RNNs) can contextualize, recognise and predict sequences of information (such as time signals and language) and have been used for numerous tasks including language, video and image processing.
There’s now a lot of interest in transferring electronic neural networks into the optical domain, creating optical neural networks that can process large data volumes at high speeds with high energy efficiency. But while there’s been much progress in general optical neural networks, work on recurrent optical neural networks is still limited.
Development of recurrent optical neural networks will require new optoelectronic devices with a short-term memory that’s programmable, computes optical inputs, minimizes noise and is scalable. In a recent study led by Birgit Stiller at the Max Planck Institute for the Science of Light, researchers demonstrated an optoacoustic recurrent operator (OREO) that meets these demands.
To read more, click here.