Fig. 6: Transformer relies on CA1 theta and BLA slow gamma to predict memory age.

a A pipeline diagram of the perturbation analysis process for the Transformer model. The red lines represent the perturbation workflow, in which a specific feature (e.g., CA1 theta) in the test samples was perturbed using a band stop filter and then input into the trained Transformer model to obtain memory age predictions without the influence of the perturbed feature. The black lines depict the workflow for testing original data, where test samples were directly fed into the trained model to obtain the actual test performance. The difference in F1 scores between these two workflows was employed to quantify the importance of the perturbed feature. b Results of the perturbation analysis for the transformer model using LFP of non-freezing periods as the input. CA1 theta and BLA slow gamma exhibited significantly higher importance compared to the baseline, which is denoted by a red line (n = 12 mice; CA1 theta, W = 78.0, p = 2 × 10⁻⁴; BLA slow gamma, W = 69.0, p = 0.008; one-tailed Wilcoxon signed-rank test). c Model for the modulation of BLA slow gamma associated with recent and remote recall. See Discussion for detail. For all bar graphs, data are represented as mean ± SEM with values from individual mice. **p < 0.01 and ***p < 0.001. Source data are available in the Source Data file.