Increased Entropic Brain Dynamics during DeepDream-Induced Altered Perceptual Phenomenology

This EEG study (n=20) appraises brain dynamics using DeepDream algorithm (to produce videos) to investigate pharmacologically-induced hallucinations. The results indicate that both DeepDream and psychedelic drugs induced similar altered brain patterns and point towards the potential of this method to study psychedelic experiences (or similar brain patterns) without giving people psychedelics.

Abstract

“In recent years, the use of psychedelic drugs to study brain dynamics has flourished due to the unique opportunity they offer to investigate the neural mechanisms of conscious perception. Unfortunately, there are many difficulties to conduct experiments on pharmacologically-induced hallucinations, especially regarding ethical and legal issues. In addition, it is difficult to isolate the neural effects of psychedelic states from other physiological effects elicited by the drug ingestion. Here, we used the DeepDream algorithm to create visual stimuli that mimic the perception of hallucinatory states. Participants were first exposed to a regular video, followed by its modified version, while recording electroencephalography (EEG). Results showed that the frontal region’s activity was characterized by a higher entropy and lower complexity during the modified video, with respect to the regular one, at different time scales. Moreover, we found an increased undirected connectivity and a greater level of entropy in functional connectivity networks elicited by the modified video. These findings suggest that DeepDream and psychedelic drugs induced similar altered brain patterns and demonstrate the potential of adopting this method to study altered perceptual phenomenology in neuroimaging research.”

Authors: Antonino Greco, Giuseppe Gallitto, Marco D’Alessandro & Clara Rastelli

Summary

1. Introduction

In recent years, there has been renewed interest in the use of psychedelics drugs to understand the neural dynamics of conscious perception. Several studies have reported increased entropy of the brain’s activity and functional connectivity, which may explain the altered states of consciousness induced by psychedelics.

To overcome ethical and legal issues, Suzuki et al. created a methodology that simulates biologically plausible visual hallucinations. This study explored the neural effects of presenting artificial hallucinations, and found that they are characterized by a more chaotic regime with respect to regular visual perception.

2.1. Participants

Twenty volunteers were recruited, with 12 females and a mean age of 26.4. They had normal vision and hearing, and gave informed written consent.

2.2. Stimuli and Procedure

Participants watched two video clips for 120 s each, one was an original clip and the other was a modified clip using DeepDream. We opted to maintain a fixed order of the conditions because participants did not recognize the modified clip as the original one. DeepDream is an algorithm that alters and enhances patterns in images through a process that can be conceived as “algorithmic pareidolia”. It relies on a pre-trained deep convolutional neural network (CNN) and uses optical flow to stabilize the optimization process.

The DD video was generated by the GoogleNet CNN with octaves, scale, jitter, zoom, step size, flow threshold, and blending ratios as described in [14].

2.3. EEG Acquisition and Preprocessing

EEG data were recorded from 27 Ag/AgCl electrodes cap at a sampling rate of 1 kHz. Spherical interpolation was performed on individual bad channels, line noise was removed, and stereotyped artifacts were deleted via independent component analysis (ICA) via the extended infomax algorithm.

2.4. Entropy and Complexity Analysis of the EEG Signal

Data were firstly analyzed in terms of entropy and complexity of the EEG signal. Permutation Entropy (PE) was used to better characterize noisy signals such as the EEG.

The elements of xk are arranged in ascending order such that xk = xk+(m1) .

A probability for each ordinal pattern can be defined as the frequency of the i-th permutation in the time series.

PE is a robust measure of nonlinear time-series, but its main limitation is its inability to differentiate between distinct forms of a certain ordinal pattern and the sensitivity of distinguishing background noise modes. A weighted version of PE was proposed to overcome this limitation.

We applied a coarse graining procedure to the time-series and calculated the multiscale weighted permutation entropy (MWPE) on the resulting signal.

We set the time scale from 1 ms to 20 ms and used the Jensen – Shannon complexity (JSC) measure to calculate the complexity of the signal.

We used a measure called multiscale weighted Jensen – Shannon complexity (MWJSC) to investigate the irregularity of the signal with respect to the entropy-periodicity plane (PE). The ECH allowed us to look at the data in a complementary way by looking simultaneously at these two aspects of the signal. We applied MWPE and MWJSC measures to the data, and also analyzed the data using multiscale versions of PE and JSC, namely MPE and MJSC. We also used Lempel – Ziv complexity (LZC) to analyze the data.

2.5. Functional Connectivity Networks and Geodesic Entropy Estimation

Functional connectivity networks were estimated from the EEG signal using a phase-based approach. The WPLI values were averaged across windows and within each frequency band, and the 95th percentile of the surrogate signals’ distribution was used to mask the actual connectivity values.

We defined the global functional connectivity (GFC) measure as the sum of all significant connections in a network, and computed the geodesic entropy (GE) measure of network entropy as the probability that one selects a node in the neighborhood of a node with geodesic distance D(i, j) = r.

The geodesic energy (GE) of a network is given by the H of the probability distribution computed on its geodesics.

2.6. Statistical Analysis

Statistical analysis was performed using cluster corrected non-parametric permutation two-sided paired t-tests on selected region of interest (ROI) and time scales. Effect size was assessed using Cohen’s d and reported as mean and max values for the significant clusters. We used linear discriminant analysis and leave-one-subject-out cross-validation to compare ECH, AGE and GFC between DD and OR conditions, and statistical significance was assessed using cluster corrected non-parametric permutation two-sided paired t-tests.

3.1. Entropy and Complexity

MWPE analysis revealed that DD videos had a higher entropy level than OR videos in the frontal ROI, and MWJSC analysis revealed a lower complexity level than OR over the parietal ROI.

The results show that entropy and complexity conjunctively differentiate the two conditions from lower to higher time scales, with a higher entropy and complexity in DD with respect to OR over frontal regions along the time scale range 6 – 16 in DD with respect to OR.

Functional connectivity analysis showed that DD had a higher gamma band GFC value and a higher average AGE than OR. The AGE was largely driven by fronto-parietal sensors.

In this paper, we investigated the brain dynamics related to artificially-induced altered perception. We found that the entropy level and the statistical complexity level were higher in the DD condition compared to the OR condition in the frontal regions, and that these differences were located at different time scales. This study found that functional connectivity networks were largely perturbed by artificial stimulation, and that the strongest increase in functional connectivity was observed in the gamma band. This finding is comparable to the results of previous studies using psychedelic drugs to induce perceptual phenomenology. The authors of DeepDream believe that the brain’s predictive process mimics the computational process underlying perceptual inference, and that the brain’s increased functional connectivity only in the gamma band could be explained by the brain’s overload of prediction error messages passing due to the highly unpredictable sensory data. The entropic brain hypothesis predicts that the brain operates at a criticality state during psychedelic experiences, and at a regime of sub-criticality during normal waking consciousness. Our findings also conform to these theoretical predictions. Future directions could include using virtual reality (VR) to present stimuli to participants while collecting EEG. A new tool to study altered perception has been developed using modern deep learning algorithms. This tool can be used to generate multiple videos that are optimized to match sequentially from low-level to high-level activation layers.