2026-08-20
Monash used CEBRA on 54 fMRI sessions: after 19 mg psilocybin, four-context SVM accuracy rose from ~62% to ~98% (r=0.64 with MEQ30); modularity fell but did not predict experience.
Psychedelic neuroimaging has spent a decade describing the brain as desynchronized and high-entropy: networks uncouple, modularity falls, time-averaged maps look messy. That account never explained the other half of the reports, the felt dissolution of self-world boundaries. Prior studies also stacked three limits: small n, experienced users, and eyes-closed rest only, often with tasks that contaminate the drug state. PsiConnect at Monash recorded 62 psychedelic-naive healthy adults on a fixed 19 mg oral dose, with fMRI about 80 minutes post-dose and EEG about 150 minutes, across rest, guided meditation, music, and an eyes-open cloud movie. The question is whether the apparent disorder hides temporal structure that tracks how deep the experience goes.
Each person's frame-by-frame BOLD from 332 parcels (Schaefer 300 cortical plus Melbourne 32 subcortical) was mapped with CEBRA-Time into a 3D trajectory. CEBRA is contrastive learning: temporally adjacent whole-brain snapshots are pulled together, dissimilar ones pushed apart, so the embedding keeps time order while making state switches visible. The four contexts were concatenated, then an SVM classified the context label at every time point. If the drug only added noise, trajectories would smear and accuracy would fall. If activity locked onto the current context, accuracy would rise.
A network perturbation test swapped one network's psilocybin time series for that same person's baseline, leaving the other networks on drug, and measured the accuracy drop. TAVRNN, a variational graph RNN with temporal attention, embedded each region's connectivity dynamics in 2D. Conventional GFC, modularity, spectral DCM, EEG power, and Lempel-Ziv complexity sat alongside as time-averaged controls. Machine-learning analyses used the 54 people who passed quality control in all four conditions.
Two design choices bound the claims. The trial was open-label with no placebo, and fMRI context order was fixed from low to high stimulation for safety on a first dose. Half the cohort sat an eight-week mindfulness course; the two arms did not differ under psilocybin and were pooled.
On time-averaged maps the drug does loosen walls. In eyes-closed conditions, sensory GFC fell by as much as 39% (Cohen's d = -0.55) while associative GFC rose by as much as 72% (d = 0.53). The eyes-open versus eyes-closed GFC gap in the visual network shrank 85% (d = -3.23). Modularity dropped in every context, d from -0.63 at rest to -0.91 during the movie. Those modularity scores did not predict MEQ30 or next-day mindset change (|r| < 0.15, all nonsignificant). Group maps move; they do not say who went deep.
Temporal embeddings reverse the picture. Four-class chance is 25%. CEBRA-SVM accuracy sat near 62% at baseline and near 98% on drug. Accuracy correlated with mean MEQ30 at r = 0.64 (P < 10^{-6}) and with next-day mindset at r = 0.40. The strongest links were positively felt self-boundary dissolution (mystical, blissful, unitive); hallucinatory subscales were weaker; anxiety and impaired control sat near zero or negative. Silhouette scores split high-MEQ from low-MEQ people only on drug (rest Δs = 0.51, meditation 0.71), not at baseline. One late-onset participant with almost no subjective effect during the scan produced a drug-day trajectory that looked like baseline.
Perturbation split the accuracy gain: default-mode +22.6 percentage points, visual +23.1, nearly half the lift; replacing either one dropped accuracy by about 6-8 points. Those two systems sit at opposite ends of the principal cortical gradient. PCA, t-SNE, and Isomap recovered context clusters too, without temporal continuity. On EEG, the eyes-open versus eyes-closed alpha-power gap fell 48% (d = -0.79). Lempel-Ziv complexity rose in eyes-closed blocks and was relatively suppressed during the movie, matching the fMRI softening of internal-external borders.
Twenty-four people ranked the session among the five most meaningful of their lives. One month later, death acceptance (d = 0.49), personal meaning (d = 0.51), and nature relatedness (d = 0.26) had moved. The eight-week mindfulness course left no group difference in connectivity or subjective scales under the drug.
For representation learning this is CEBRA moved from electrophysiology onto whole-brain fMRI, with a subjective-scale readout attached. The temporal structure contrastive learning keeps is exactly what modularity throws away. Accuracy going up on drug means looking messier can sit next to separating more cleanly by context. For consciousness and psychiatric imaging, embeddedness (their name for feeling continuous with the environment rather than separate from it) turns a hard-to-operationalize self-boundary report into trajectory geometry a classifier can score. Data are public on OpenNeuro as ds006110.
No new architecture. The move is to put existing embeddings on a large, multi-context acute dataset. Incremental as a method paper, and it does split the usual line that psychedelics equal entropy and disorder.
Open-label, no active placebo, unblinded even at analysis, so expectancy is not washed out. fMRI context order was fixed; rest already separated high-MEQ clusters as the first block, but order confounds remain. EEG started around 150 minutes post-dose, after the peak. The sample is healthy and psychedelic-naive, so it does not travel to depression or addiction treatment. The ML slice is n = 54 with no independent cohort; the authors say generalization across populations is untested. The next-day mindset instrument is homemade, and subscale brain-behavior tests were uncorrected. Embeddedness is a post-hoc construct; correlation is not mechanism. A null mindfulness arm only says this course at this dose did not move the needle.