MinedDeep

Semantic Representations in the Temporal Pole Predict False Memories

3 min read

Semantic Representations in the Temporal Pole Predict False Memories

Type: paper Slug: semantic-representations-in-the-temporal-pole-predict-false-memories—hassabis Sources: semantic-representations-in-the-temporal-pole-predict-false-memories—hassabis Last updated: 2026-05-13

Summary

Chadwick, Anjum, Kumaran, Schacter, Spiers, and Hassabis (2016) used fMRI with representational similarity analysis (RSA) to show that a similarity-based neural code in the left temporal pole predicts DRM false memory likelihood. Each individual’s unique temporal pole representations predicted their idiosyncratic false memory errors, and the same code also predicted true memory performance — consistent with an adaptive view where semantic knowledge both enhances and distorts memory.

Core content

Research question: What is the neural basis of semantic false memory — specifically, does the temporal pole contain a similarity-based code that can predict false memory formation?

Design: 18 participants viewed 40 four-word DRM lists and 40 related lure words during fMRI (incidental semantic categorization task). A separate behavioral session (weeks earlier) measured individual false memory patterns. Searchlight RSA measured neural overlap between each lure and its list items, then correlated this with canonical false memory rates (paper—semantic-representations-in-the-temporal-pole-predict-false-memories §Results).

Key findings:

  • Whole-brain searchlight revealed a single significant cluster in the left temporal pole (peak: −51, 17, −25; Pseudo-t=5.33; 92 voxels) where neural overlap correlated with false memory likelihood across the 40 DRM lists (paper—semantic-representations-in-the-temporal-pole-predict-false-memories §Results).
  • Cross-validated ROI analysis confirmed the correlation (r(39)=0.40, p=0.012), robust after removing outliers (r=0.45, p=0.005) (paper—semantic-representations-in-the-temporal-pole-predict-false-memories §Results).
  • Individuation analysis: within-subject neural-behavioral correlations significantly exceeded between-subject correlations (Z=2.63, p=0.004), showing each person has a partially unique semantic code that predicts their own false memory errors (paper—semantic-representations-in-the-temporal-pole-predict-false-memories §Results).
  • The same TP code predicted true memory strength (Z=2.33, p=0.009), including individual differences (Z=2.16, p=0.016) (paper—semantic-representations-in-the-temporal-pole-predict-false-memories §Results).
  • Effects survived controls for word frequency, visual similarity (Levenshtein distance), and task-driven categorical representations (paper—semantic-representations-in-the-temporal-pole-predict-false-memories §Results).
  • 1.5T scanner was deliberately chosen over higher-field options to reduce temporal pole susceptibility artifacts (paper—semantic-representations-in-the-temporal-pole-predict-false-memories §Methods).

Theoretical contribution: Provides the first direct neural evidence that computational models positing a similarity-based semantic code in the temporal pole (Patterson et al., 2007; McClelland & Rogers, 2003) are correct, and that this code has a behavioral cost — false memories — as an emergent property.

Connections- Theme: theme—hippocampal-construction, episodic-memory

  • Collaborators: Martin J. Chadwick (first author), Raeesa S. Anjum, Dharshan Kumaran, Daniel L. Schacter, Hugo J. Spiers (joint senior with Hassabis)
  • Era: deepmind-era — Hassabis listed at Google DeepMind
  • Venue: venue—PNAS (NOT Nature Neuroscience as metadata states)
  • Influences: Patterson et al. (2007) semantic hub model; Schacter et al. (2011) adaptive memory distortion
  • Notable quote: “Our findings reveal the underlying structure of neural representations of semantic knowledge, and how this semantic structure can both enhance and distort our memories.” (paper—semantic-representations-in-the-temporal-pole-predict-false-memories §Abstract)

Honest Gaps

  • Metadata incorrectly lists venue as Nature Neuroscience; the paper was published in PNAS.
  • Only 4 of the standard 15 DRM list items were used due to scanning time constraints, potentially reducing false memory effect sizes.
  • Behavioral and fMRI sessions were separated by 21–239 days (mean 65 days); the long delay was intentional but introduces noise.
  • The study cannot determine whether TP-hippocampal interactions generate false memories, only that TP codes the semantic similarity that predicts them.
  • 18 participants is modest for an individuation analysis.

Graph View

Backlinks