pre-trained word vectors from the word2vec model  (Baidu Inc)

Journal: Nature Communications

Article Title: Insight predicts subsequent memory via cortical representational change and hippocampal activity

doi: 10.1038/s41467-025-59355-4

Figure Lengend Snippet: Insight was analyzed as a continuous variable but for visualization purposes, it was median-split into low, (medium) and high insight values. Asterisk = statistical significance at p < 0.05; A RC from pre to post solution: multivoxel pattern similarity. Multivoxel patterns per ROI for each time point (pre and post solution) are extracted and subsequently correlated. Pre = 0.5 s after stimulus presentation; post = during solution button press. Those Pre-Post Solution Similarity values ( r ) are subsequently estimated in a linear mixed model as a function of insight and ROI. Bar plots show change (Δ) in Multivoxel Pattern Similarity (MVPS) analysis. Change in MVPS = 1 minus the correlation between the post and pre solution multivoxel pattern in the respective ROI. Violin plots represent predicted data ( n = 19,536 samples) from single trial analysis for 6 bilateral VOTC regions; points indicate estimated marginal means; error bars represent 95% (between-subject) confidence intervals. Only the estimated slope coefficients (two-sided t -test, with Bonferroni correction for multiple comparisons) of bilateral iLOC ( t (19,514.5) = −8.08, p < 0.001, ß = 0.02, 95% CI [0.01, 0.02]) and pFusG ( t (19,514.5) = −6.09, p < 0.001, ß = 0.01, 95% CI [0.01, 0.01]) showed significant changes in MVPS in the hypothesized direction. B RC from pre to post solution: Representational strength—AlexNet. This RSA method employs four steps. (1) A neural activity-derived RSM (Brain RSM, size 120 × 120) is generated for each region-of-interest (ROI) and each time point (pre post solution, see A ) where each cell is representing a multivoxel pattern similarity value for each Mooney image pair. (2) A conceptual Model RSM (here using AlexNet, size 120 × 120) is generated where each cell is representing a similarity value for each Mooney object pair. (3) For each brain region and each time point, the row of each stimulus (~120) in the Model RSM and in the Brain RSM are correlated yielding a representational strength measure (i.e., brain-model fit) per region and time point. (4) The representational strength is used as a dependent variable in linear mixed models to investigate which ROIs exhibit an insight-related increase in representational strength from pre to post solution (time). Violin plots depict predicted Representational Strength (Rep-Str) values at single-trial level ( n = 14,652 samples), defined as the second-order correlation between multivoxel patterns in the respective ROI at pre and post-response time points and conceptual Model RSMs (AlexNet and Word2Vec [W2V]); points indicate estimated marginal means ± SEM. Bilateral iLOC (AlexNet: Chi ²(1) = 6.70, p = 0.010, ß = 0.07, 95% CI [0.02, 0.12]; W2V: Chi ²(1) = 12.39, p < 0.001, ß = 0.07, 95% CI [0.03, 0.10]) and pFusG (AlexNet: Chi²(1) = 8.69, p = 0.003, ß = 0.06, 95% CI [0.02, 0.09]; W2V: Chi ²(1) = 3.92, p = 0.048, ß = 0.05, 95% CI [0.00, 0.11]) show a significant increase in Rep-Str over time, based on two-sided Likelihood ratio tests (uncorrected for multiple comparison). The brain images were generated using MRIcroGL .

Article Snippet: D Results display combined representational strength of solution object for iLOC & pFusG measured via a conceptual Word2Vec [W2V] model. Asterisk represents Insight*Memory*Time interaction at p < 0.05 ( Chi 2(2) = 10.46, p = 0.004, ß = 0.09, 95% CI [0.02, 0.15], n = 13,860 samples).

Techniques: Activity Assay, Derivative Assay, Generated, Comparison

Journal: Nature Communications

Article Title: Insight predicts subsequent memory via cortical representational change and hippocampal activity

doi: 10.1038/s41467-025-59355-4

Figure Lengend Snippet: Amy Amygdala, aHC anterior hippocampus, pHC posterior hippocampus, pFusG posterior Fusiform Gyrus, iLOC inferior Lateral Occipital Lobe. Violin plots illustrate the predicted estimates at a single-trial level, with dots representing the estimated marginal means ± SEM (likelihood ratio tests, uncorrected for multiple comparison). Insight was analyzed as a continuous variable but for visualization purposes, it was split into low, medium and high insight values. A Change (Δ) in Multivoxel Pattern Similarity (MVPS). Change in MVPS = 1 minus the correlation between the post and pre solution multivoxel pattern in the respective ROI. Asterisk represents Insight*Memory interaction for iLOC ( Chi ²(1) = 4.08, p = 0.038, ß = −0.07, 95% CI [−0.14, −0.00], n = 2310 samples, tested two-sided) and pFusG ( Chi ²(1) = 5.84, p = 0.014, ß = −0.07, 95% CI [−0.12, −0.01]), n = 4620 samples, tested two-sided). B Results display combined representational strength of solution object for iLOC & pFusG measured via a conceptual model created out of the penultimate layer of AlexNet. There was no statistically significant Insight*Memory*Time effect ( p > 0.40; n = 13,860 samples). C Violin plots represent beta estimates of BOLD activity at post solution divided by insight memory conditions. Asterisk represents Insight*Memory interaction for aHC ( Chi² (1) = 5.68, p = 0.015, ß = 0.09, 95% CI [0.02, 0.17], n = 2976 samples). D Results display combined representational strength of solution object for iLOC & pFusG measured via a conceptual Word2Vec [W2V] model. Asterisk represents Insight*Memory*Time interaction at p < 0.05 ( Chi ²(2) = 10.46, p = 0.004, ß = 0.09, 95% CI [0.02, 0.15], n = 13,860 samples). The brain image was generated using CONN [RRID:SCR_009550].

Article Snippet: D Results display combined representational strength of solution object for iLOC & pFusG measured via a conceptual Word2Vec [W2V] model. Asterisk represents Insight*Memory*Time interaction at p < 0.05 ( Chi 2(2) = 10.46, p = 0.004, ß = 0.09, 95% CI [0.02, 0.15], n = 13,860 samples).

Techniques: Comparison, Activity Assay, Generated