A Scheme for Optical Reservoir Computers with Atomic Memory

Optical systems have been identified as one of the promising alternative hardware implementations for machine learning applications due to low latency, high bandwidth and lower energy consumption from the use of passive optics [1] . Reservoir computers (RCs) are interesting as they are versatile and easily trained recurrent neural networks - preforming particularly well on time series data [2] . RCs lend themselves particularly well an optics implementation, as the the reservoirs remained fixed, and the trained weights exchanged, allowing for reuse of one reservoir for multiple problems. It has been demonstrated that optic-to-electric conversion introduces uncertainty and time delay in the implementation of optical reservoir computers [3] , thus motivating research into all-optical RCs. First implementations of an all-optical RC used feedback loop laser setups with fixed delay and dynamics, limiting state space which can be represented in the reservoir, but allowing quick data throughput [2] , [4] . We propose an all-optical setup with the internal states of the reservoir which are stored in an optical memory, allowing for more complex internal dynamics. We present an electro-optical implementation as first result, and outline how this can be implemented all-optically by using atomic vapor based memories [5] .