Extended Data Fig. 5: Convergence analysis Ecomplex STDP and Double T-Maze experiment.

The ENU can be seen to generally converge faster in both experimental settings. In case of evolving a complex synaptic update rule (Complex STDP), the ENU significantly outperforms the other models. When the feedback connection is removed (NFENU), the performance also drops. This indicates the importance of the feedback connection, which was also observed in the previous standard STDP experiment in Fig. 4. In case of the Double T-Maze experiment, the ENU also converges faster with this feedback connection. The LSTM generally takes longer to converge compared to the GRU model, which could be explained by the fact that LSTMs are slightly more complex than GRUs. When the parameters are shared between the synapse and neuron ENUs, the network fails to converge (SHAREDENU). This was also observed in the standard T-Maze experiment, and further indicates the need for the specialization of the synaptic and neuronal behaviour.