These results mostly support the novel ensemble model. For much of the post-saccade delay, decoders generalized poorly from the constant contralateral trials to those in which it was transferred from the opposite hemisphere (Figure 6E). The initial bump in decoding accuracy seen when the saccade shifted the remembered location from ipsilateral to contralateral (Figure 4B) was not detected using the classifiers from no-swap trials. Thus, different neural ensembles were activated by the same sensory information in WM, depending on whether it arrived via ipsilateral visual inputs or via the contralateral hemisphere. However, near the end of the delay, the cross-classification of contralateral-shift trials increased relative to the control (Figure 6E, p < 0.01, paired t test). This suggests that, in anticipation of using the WM, interhemispherically transferred memory traces converged somewhat toward an ensemble representation similar to feedforward-induced traces.
