Figure 3.

The network inherits major generic processing capabilities of the liquid computing model, as shown in Figure 12. A, In the XOR task, a readout neuron has to decide whether specific combinations of input patterns are currently presented to the network. It should respond strongly if pattern A at input stream 1 is accompanied by pattern B′ in stream 2 or pattern A′ by pattern B, but not respond for other combinations of the same patterns. B, Samples of the four spike patterns A, A′, B, and B′. Each pattern has 50 ms duration and consists of Poisson spike trains at 5 Hz. C, Sample stimulus sequence with the patterns shown in B. During simulation, 2 Hz Poisson spike trains are overlaid as noise (black spikes). In the XOR task, the readout should respond only if one of the desired pattern combinations is currently present at the input (indicated by blue rectangles). In a separate memory task, the readout should decode the identity of the pattern presented during the interval [−100 ms, −50 ms] in stream 2 (magenta rectangles). D, Performance of a linear regression readout for both tasks trained on the low-pass-filtered spike trains of both the network and the stimulus directly (time constant 20 ms). The readout was trained to decode the target bit every 50 ms at the end of each pattern (training set 50 s/1000 patterns); shown is the performance on a test set (6 s/120 patterns), averaged over 100 runs with different patterns and networks (error bars show SD). Performance is evaluated as the point-biserial correlation coefficient between the binary target variable and the analog output of the linear regression.