Temporal Information Encoding in Isolated Cortical Networks

Abstract Time-dependent features are present in many sensory stimuli. In the sensory cortices, timing features of stimuli are represented by spatial as well as temporal code. A potential mechanism by which cortical neuronal networks perform temporal-to-spatial conversion is ‘reservoir computing’. The state of a recurrently-connected network (reservoir) represents not only the current stimulus, or input, but also prior inputs. In this experimental study, we determined whether the state of an isolated cortical network could be used to accurately determine the timing of occurrence of an input pattern – or, in other words, to convert temporal input features into spatial state of the network. We used an experimental system based on patterned optogenetic stimulation of dissociated primary rat cortical cultures, and read out activity via fluorescent calcium indicator. We delivered input sequences of patterns such that a pattern of interest occurred at different times. We developed a readout function for network state based on a support vector machine (SVM) with recursive feature elimination and custom error correcting output code. We found that the state of these experimental networks contained information about inputs for at least 900 msec. Timing of input pattern occurrence was determined with 100 msec precision. Accurate classification required many neurons, suggesting that timing information was encoded via population code. Trajectory of network state was largely determined by spatial features of the stimulus, with temporal features having a more subtle effect. Local reservoir computation may be a plausible mechanism for temporal/spatial code conversion that occurs in sensory cortices. Significance Statement Handling of temporal and spatial stimulus features is fundamental to the ability of sensory cortices to process information. Reservoir computation has been proposed as a mechanism for temporal-to-spatial conversion that occurs in the sensory cortices. Furthermore, reservoirs of biological, living neurons have been proposed as building blocks for machine learning applications such as speech recognition and other time-series processing. In this work, we demonstrated that living neuron reservoirs, composed of recurrently connected cortical neurons, can carry out temporal-spatial conversion with sufficient accuracy and at sufficiently long time scale to be a plausible model for information processing in sensory cortices, and to have potential computational applications.