Back to experiment

Brain Region Analysis — Real vs. Reversed Physics

86.0% CV accuracy, 0.946 AUC; 95% holdout accuracy

The classifier was 99.0% sparse — only 202 of 20,484 vertices had non-zero weights. This experiment used video input (V-JEPA2 encoder) rather than text, making the results free from TTS pipeline artifacts. Weights were mapped onto the Destrieux cortical atlas (fsaverage5).

Real-physics-predictive regions

Regions whose activation predicts real (forward) physics videos (positive classifier weights).

RH Cingulate Cortex (mid-posterior + dorsal + marginal)

~23.7%Strong

This is the dominant real-physics-predictive region. The anterior and mid-cingulate cortex (ACC/MCC) functions as a forward model that predicts future states (Alexander & Brown, 2019). Limongi et al. (2013) showed that temporal prediction errors — the discrepancy between when an event was expected to occur and when it actually occurred — modulate coupling between the cingulate and insula. For real physics videos, the cingulate can successfully build a forward model: a ball falls, accelerates under gravity, bounces. In reversed videos, the forward model breaks down — objects spontaneously launch upward, fluids reassemble — which is why the cingulate activates more for real physics where predictions are coherent and sustained.

Makes strong neuroscientific sense. The cingulate as a forward model for causal event prediction is well-documented.

RH Superior Frontal Gyrus (mPFC)

5.0%Strong

Part of the mPFC, involved in predictive coding and maintaining internal models. Fischer et al. (2016) identified a network including premotor cortex and supplementary motor area that activates during intuitive physical reasoning. The mPFC likely supports high-level prediction and model-updating when watching coherent physical scenes.

Makes strong neuroscientific sense. mPFC is involved in predictive processing and internal model maintenance.

Bilateral Posterior Lateral Fissure

7.6%Strong

The posterior lateral fissure borders the superior temporal gyrus and the inferior parietal lobule, near the temporoparietal junction. This region is adjacent to motion-sensitive areas (MT+/V5) and areas involved in event structure processing. It would process the temporal dynamics and causal structure of physical events — when objects move in physically consistent ways, this region can extract meaningful event structure.

Moderately to strongly supported. Temporal/parietal junction regions process event structure and motion.

Bilateral + LH Inferior Temporal Sulcus + Middle Temporal Gyrus

~10.7%Strong

These temporal regions are involved in higher-order visual processing, including motion perception, object recognition, and event structure. The middle temporal gyrus (near MT+/V5) specifically processes visual motion. Real physics videos have coherent, physically plausible motion patterns that this region would process more efficiently.

Makes strong neuroscientific sense. Temporal motion-processing regions preferentially encode coherent physical motion.

Reversed-physics-predictive regions

Regions whose activation predicts time-reversed physics videos (negative classifier weights). These are the brain's "error detectors."

LH Intraparietal Sulcus (IPS)

8.2%Strong

This is perhaps the most important finding. Schwettmann et al. (2019) showed that dorsal fronto-parietal cortex — including the IPS — encodes invariant representations of physical variables like mass, generalizing across different scenarios, materials, and motion patterns. Fischer et al. (2016) identified a broader "physics engine" network (primarily premotor cortex and supplementary motor area) that activates during intuitive physical reasoning. The IPS is part of this extended network. Here, the IPS predicts reversed physics. This is counterintuitive at first, but has a compelling explanation: when watching reversed videos, the brain's physics engine detects violations — objects moving in physically impossible ways — and activates more intensely as it tries to reconcile the visual input with its internal physical model. This is consistent with prediction error theory.

Makes strong neuroscientific sense. The IPS (the brain's "physics engine") firing for reversed physics is consistent with prediction error — it fires harder when physics is wrong.

LH Insula (inferior circular + short gyri + central)

~15.6%Strong

The insula is the brain's primary salience and prediction error hub. The anterior insula encodes risk prediction errors (Bossaerts, 2010) and signals deviations from expectations via bursts of beta oscillations (Haufler et al., 2022). Limongi et al. (2013) showed that temporal prediction errors specifically modulate cingulate-insular coupling. Reversed physics videos represent massive violations of temporal and physical expectations. Objects move at the wrong times, in the wrong directions, with impossible causal sequences.

Makes strong neuroscientific sense. The insula as a prediction-error detector explains its strong reversed-physics activation perfectly.

LH Parahippocampal Gyrus

8.5%Moderate

The parahippocampal gyrus encodes spatial layout and contextual associations (Epstein & Kanwisher, 1998). Research shows it is primarily engaged when processing scenes and spatial context — it responds strongly to congruent, meaningful scene contexts. Our interpretation is that reversed physics videos disrupt normal contextual processing: objects appear in impossible spatial configurations (liquid flying upward, objects launching off surfaces), which would alter the parahippocampal response compared to normal physics. However, calling it a "violation detector" would be an overstatement — its established function is encoding contextual associations, not detecting violations per se.

The parahippocampal gyrus's role in contextual encoding is well-established. That it responds differently to reversed physics is our interpretation — plausible but not directly tested in the literature.

LH Supramarginal Gyrus (TPJ)

4.1%Strong

Part of the temporoparietal junction (TPJ), the supramarginal gyrus is involved in temporal order judgments. Davis, Christie & Rorden (2009) showed that the TPJ activates specifically when making judgments about the temporal order of events. Reversed videos inherently violate temporal order — events happen backward — which would specifically engage this temporal-order processing region.

Makes strong neuroscientific sense. The supramarginal gyrus processes temporal order, which is violated in reversed videos.

RH Frontopolar Gyrus

6.7%Moderate

The frontopolar cortex (Brodmann area 10) is involved in monitoring and evaluating competing cognitive representations. When watching reversed physics, the brain simultaneously processes the visual input (what it sees) and the expected physics (what should happen), creating a conflict that engages the frontopolar cortex's conflict-monitoring function.

Moderately supported. Frontopolar conflict monitoring is plausible but less directly established for physics perception.

Overall verdict

The physics experiment's brain regions tell a coherent story about how the brain processes physical plausibility. Real physics activates regions associated with forward modeling — the cingulate cortex (which predicts future states), temporal motion areas (which process coherent motion), and the mPFC (which maintains internal models). Reversed physics activates regions associated with error detection — the insula (salience/prediction error), the intraparietal sulcus (physical variable encoding, responding to violations), the parahippocampal gyrus (contextual encoding disrupted by impossible spatial configurations), and the supramarginal gyrus (temporal order processing). The prediction-vs-error framing is our interpretation of the pattern, but each individual region's known function aligns with the direction of the finding.

References

  1. 1.Alexander, W. H., & Brown, J. W. (2019). The role of the anterior cingulate cortex in prediction error and signaling surprise. Topics in Cognitive Science, 11(1), 106-120.
  2. 2.Bossaerts, P. (2010). Risk and risk prediction error signals in anterior insula. Brain Structure and Function, 214, 645-653.
  3. 3.Davis, B., Christie, J., & Rorden, C. (2009). Temporal order judgments activate temporal parietal junction. Journal of Neuroscience, 29(10), 3182-3188.
  4. 4.Epstein, R., & Kanwisher, N. (1998). A cortical representation of the local visual environment. Nature, 392, 598-601.
  5. 5.Fischer, J., et al. (2016). Functional neuroanatomy of intuitive physical inference. PNAS, 113(34), E5072-E5081.
  6. 6.Haufler, A. J., et al. (2022). Human anterior insula signals salience and deviations from expectations via bursts of beta oscillations. Journal of Neurophysiology, 128(1), 160-176.
  7. 7.Limongi, R., et al. (2013). Temporal prediction errors modulate cingulate-insular coupling. NeuroImage, 71, 147-157.
  8. 8.Schwettmann, S., et al. (2019). Invariant representations of mass in the human brain. eLife, 8, e46619.

Analysis based on Destrieux cortical atlas (fsaverage5).

Back to experiment