Increased Entropic Brain Dynamics during DeepDream-Induced Altered Perceptual Phenomenology

Greco and colleagues situate this study within a resurgence of neuroscientific interest in psychedelic substances (for example LSD, psilocybin, ketamine) as tools to probe the neural dynamics of conscious perception. Earlier empirical work has reported increased signal entropy, a larger repertoire of dynamical brain states and altered functional connectivity during pharmacologically induced psychedelic states. Those findings underpin the entropic brain hypothesis, which characterises psychedelic altered consciousness as a higher-entropy regime of brain activity. However, the authors note practical and ethical impediments to pharmacological studies—legal restrictions and difficulty separating direct neural effects from systemic physiological consequences of the drugs. To address these obstacles, the investigators adapted a non-pharmacological method to simulate hallucination-like visual input using the DeepDream algorithm, building on prior behavioural work (the Hallucination Machine). The study set out to test whether DeepDream-modified videos elicit EEG signatures akin to those reported for psychedelic states. Specifically, Greco sought to compare entropy and statistical complexity of EEG time series across multiple time scales, and to examine entropy-related measures of functional connectivity networks, contrasting responses to a regular video (OR) and its DeepDream-modified counterpart (DD). This is presented as the first neurophysiological investigation of artificially induced altered perceptual phenomenology using this approach.

Twenty healthy volunteers (12 female, mean age 26.4 years) participated; all were right-handed, had normal or corrected vision and hearing, reported no neurological disorders and were not taking neurological medications. The extracted text does not clearly report the age range. Participants provided written consent and the protocol received local ethics approval. Stimuli consisted of two 120 s video clips displayed on a CRT monitor at 1 m distance: an original clip (OR) and a DeepDream-transformed version (DD). The DD video was generated by applying the DeepDream algorithm to the original clip, adapting the static-image algorithm to video by incorporating optical flow to stabilise hallucination-like patterns across frames. A higher-level layer (inception_4d/pool) of a pre-trained GoogleNet CNN was targeted and the authors report using hyperparameters aligned with prior work (e.g. multiple octaves, iterative gradient ascent and optical-flow blending). The order of presentation was fixed (OR then DD) to avoid potential carry-over effects from the strong DD experience; the team judged recognition of the modified clip to be low, reducing concerns about order-based learning. EEG was recorded with a 27-electrode montage based on the 10–5 system, sampled at 1 kHz, referenced to the right mastoid and grounded at AFz. Preprocessing included interpolation of bad channels (average 1.4 channels interpolated), band-pass filtering (0.1–80 Hz), line-noise removal, re-referencing to the common average and artifact removal by ICA (extended infomax), with an average of 11.95 independent components removed. Processed data were converted to FieldTrip format for analysis. Time-series analyses focussed on multiscale information measures. For entropy the authors used permutation entropy (PE) and a weighted variant (WPE) that incorporates amplitude information, embedding with dimension m=5 and delay τ=1. They applied a coarse-graining procedure across time scales θ = 1…20 ms and computed multiscale weighted permutation entropy (MWPE). Complexity was quantified using a Jensen–Shannon complexity measure adapted to the weighted entropy (MWJSC); the entropy–complexity hyperplane (ECH) combined these dimensions. For completeness, classical multiscale PE and JSC (MPE, MJSC) and two variants of Lempel–Ziv complexity (median binarised mLZC and permutation-based pLZC) were also computed. Functional connectivity networks were estimated from 1 s non-overlapping windows by computing cross-spectral density and the weighted phase lag index (WPLI) in alpha (7–13 Hz), beta (13–25 Hz) and gamma bands. Significance of connections was assessed via 200 phase-randomised surrogates and thresholding at the 95th percentile, yielding binarised networks. Global functional connectivity (GFC) was defined as the count of significant connections; geodesic entropy (GE), a network-level entropy metric based on shortest-path distributions around each node, was averaged across nodes to obtain AGE. Statistical testing used cluster-corrected non-parametric permutation paired t-tests (two-sided, 10,000 randomisations, α = 0.05). ROIs were defined for frontal, temporo-parietal and occipital sensor groups and clusters were evaluated across space and time-scale or frequency dimensions as appropriate. Effect sizes were reported with Cohen’s d. The ECH representation was also analysed with multivariate pattern analysis (linear discriminant analysis), using leave-one-subject-out cross-validation and area under the curve (AUC) as the performance metric; statistical significance was assessed via label permutation tests.

Entropy and complexity analyses of single-channel EEG showed complementary effects in frontal sensors. MWPE indicated higher entropy for the DD compared with OR in the frontal ROI across time scales θ = 7–17 ms (p = 0.036, mean d = 0.60, max d = 0.71). Topographic maps showed a broader spatial trend of increased MWPE for DD versus OR. In contrast, MWJSC (multiscale weighted Jensen–Shannon complexity) revealed lower statistical complexity for DD in the frontal ROI across time scales θ = 11–19 ms (p = 0.043, mean d = 0.58, max d = 0.71). The authors report that these effects occupied different time-scale ranges: entropy increases at lower time scales and complexity decreases at higher time scales. Analyses using the unweighted multiscale measures produced similar patterns but with smaller effect sizes: MPE showed higher entropy in frontal regions for DD over θ = 6–16 ms (p = 0.033, mean d = 0.59, max d = 0.63), and MJSC showed decreased complexity for DD over θ = 7–17 ms (p = 0.027, mean d = 0.59, max d = 0.64). Lempel–Ziv measures were less sensitive: median-binarised LZC (mLZC) showed no significant differences, while permutation-LZC (pLZC) displayed a non-significant trend of increased DD entropy over frontal regions at θ = 9–14 ms. Multivariate classification on the entropy–complexity hyperplane corroborated these effects: two time-scale clusters yielded above-chance discrimination between conditions, cluster θ = 6–11 ms (p = 0.008, mean AUC = 0.80) and cluster θ = 18–20 ms (p = 0.017, mean AUC = 0.69). Functional connectivity analyses revealed increased global functional connectivity (GFC) for DD relative to OR selectively in the gamma band, with GFC higher in DD (p = 0.016, d = 0.59). Geodesic entropy averaged across nodes (AGE) was also higher for DD in the gamma band (p = 0.041, d = 0.51), and sensor-level GE maps suggested the effect was driven largely by fronto-parietal sensors. The authors present these network effects as a greater number of significant gamma-band connections and higher network-level entropy during DD.

Greco interprets the main findings as evidence that DeepDream-induced visual stimulation evokes brain dynamics resembling those reported in pharmacological psychedelic states. The pattern reported is an increase in EEG entropy and a concomitant decrease in statistical complexity in frontal regions, occurring at different temporal scales, together with increased gamma-band functional connectivity and higher network geodesic entropy. The authors relate these results to prior psychedelic research, noting parallels with studies reporting frontal modulation and increased between-network connectivity under substances such as ayahuasca and psilocybin. They argue that the multiscale approach and the use of weighted entropy and complexity measures were important to reveal the spatiotemporal structure distinguishing normal and altered perception; by contrast, LZC appeared less sensitive in this paradigm. Two theoretical perspectives are offered to account for the results. First, within predictive coding, DeepDream can be seen as imposing a strong perceptual prior that generates large prediction errors; increased gamma-band connectivity is interpreted as reflecting an overload of prediction-error signalling, which is typically conveyed at higher frequencies. Second, the findings are discussed in the light of the entropic brain hypothesis: simulated hallucinations appear to push the brain toward a higher-entropy, more critical regime with access to a broader repertoire of states. Greco and colleagues propose future directions including presentation of DeepDream stimuli via immersive virtual reality to increase ecological validity, systematic exploration of DeepDream parameter space (for example varying targeted CNN layers to move from low-level to high-level feature modulation), and direct comparisons between DeepDream-evoked states and pharmacological psychedelic experiences. They also suggest examining additional neural signatures such as alpha suppression and directed connectivity. The authors conclude that modern deep learning algorithms offer a promising, non-pharmacological tool for studying altered perceptual phenomenology, while noting the need for further work to define the degree of comparability with true psychedelic states.