Fig. 6: Comparing optimized model behavior under ablation of the attention mechanism with recall behavior of medial temporal lobe (MTL) patients with amnesia.

A–C Behavioral patterns of patients with MTL amnesia (N = 10) compared to healthy controls (N = 16), demonstrating memory deficits on a free recall task, reproduced from Palombo et al.³⁶. Despite the small list size, patients with MTL amnesia exhibit diminished recall performance and reduced ability to perform backward recall transitions (conditional response probability at −1 lag). D–F Behavioral patterns of the seq2seq model with and without attention (hidden dimension size 64), optimized or during intermediate training. The model without attention exhibits diminished recall performance and reduced ability to perform backward recall transitions (no backward contiguity at any stage of training as seen from the conditional response probability at −1 lag). G, H Behavioral patterns of models of varying hidden dimension sizes with and without attention. Models without attention can achieve the optimal recall behavior but require sufficiently large hidden dimensions. (N = 13 for each dimension size evaluation).