Abstract
Reservoir computation is a recurrent framework for learning and predictingtime series data, that benefits from extremely simple training andinterpretability, often as the the dynamics of a physical system. In thispaper, we will study the impact of noise on the learning capabilities of analogreservoir computers. Recent work on reservoir computation has shown that theinformation processing capacity (IPC) is a useful metric for quantifying thedegradation of the performance due to noise. We further this analysis anddemonstrate that this degradation of the IPC limits the possible features thatcan be meaningfully constructed in an analog reservoir computing setting. Weborrow a result from quantum complexity theory that relates the circuit modelof computation to a continuous time model, and demonstrate an exponentialreduction in the accessible volume of reservoir configurations. We conclude byrelating this degradation in the IPC to the fat-shattering dimension of afamily of functions describing the reservoir dynamics, which allows us toexpress our result in terms of a classification task. We conclude that anyphysical, analog reservoir computer that is exposed to noise can only be usedto perform a polynomial amount of learning, despite the exponentially largelatent space, even with an exponential amount of post-processing.