Extended Data Fig. 4: Inverted encoding modeling (IEM) also reveals item-specific information represented in MTL high-frequency (70–150 Hz) activity during STM retention.

(a) The IEM method assumes that multivariate activity in a region (B) reflects a weighted summation of ideal stimulus information channels (C) of recorded neural populations (top right)⁴³. The weights (W) of these information channels can be learned from brain activity in a training dataset (top left) and then applied to an independent hold-out test dataset to reconstruct stimulus information channels (C’). The resultant vector length of the normalized channel response function can be interpreted as a measure of STM precision, when the probability of recall success is high (bottom right)³⁹. (b) Based on this approach, we have replicated observations from the more assumption-free representational similarity analysis approach. That is, information related to the cued item emerges following stimulus offset and lasts for about 500 ms during the retention interval. In contrast, information related to the uncued item is minimal and is significantly weaker as compared with that related to the cued item during STM retention interval. a.u. = arbitrary unit. Significant timepoints after cluster-based correction for multiple comparisons at the 0.05 level are marked in yellow.