Temporal irreversibility of neural dynamics as a signature of consciousness

Abstract

Dissipative systems evolve in the preferred temporal direction indicated by the thermodynamic arrow of time. The fundamental nature of this temporal asymmetry led us to hypothesize its presence in the neural activity evoked by conscious perception of the physical world, and thus its covariance with the level of conscious awareness. We implemented a data-driven deep learning framework to decode the temporal inversion of electrocorticography signals acquired from non-human primates. Brain activity time series recorded during conscious wakefulness could be distinguished from their inverted counterparts with high accuracy, both using frequency and phase information. However, classification accuracy was reduced for data acquired during deep sleep and under ketamine-induced anesthesia; moreover, the predictions obtained from multiple independent neural networks were less consistent for sleep and anesthesia than for conscious wakefulness. Finally, the analysis of feature importance scores highlighted transitions between slow ($\approx$20 Hz) and fast frequencies (>40 Hz) as the main contributors to the temporal asymmetry observed during conscious wakefulness. Our results show that a preferred temporal direction is manifest in the neural activity evoked by conscious mentation and in the phenomenology of the passage of time, establishing common ground to tackle the relationship between brain and subjective experience.

Nonequilibrium thermodynamics of the asymmetric Sherrington-Kirkpatrick model

“…p(s u , s u−1 ) log p(s u |s u−1 ), (6) which at the steady state corresponds to the Kolmogorov-Sinai entropy or entropy rate, lim t→∞ In stochastic processes in thermodynamic equilibrium, forward and time-reversed trajectories are indistinguishable. Markov processes can break this time-reversal symmetry, capturing the irreversible characteristics of physical and biological processes [38].…”

“…Recently, nonequilibrium properties of biological and intelligent systems have received the attention of neuroscience and biological science communities. For example, increased entropy production in macroscopic neural activity was suggested as a signature of physically and cognitively demanding tasks [4], conscious activity [5,6] or neuropsychiatric diseases like schizophrenia, bipolar disorders, and ADHD [7]. While it is not easy to contrast their findings based on the coarse-grained analysis of ECoG or fMRI data with the present results, our precise characterization of the entropy production of the pro-totypical system sheds light on what kind of behaviours we might expect from these complicated systems.…”

“…Understanding these dissipative processes -describing spatial and temporal patterns with a definite past-future order, e.g., chemical reactions, neural dynamics, or flocks of birds-brings critical insights into how open systems self-organize [2]. Although these ideas have attracted the interest of disparate fields from evolutionary dynamics [3] to neuroscience [4][5][6][7], little is known about the thermodynamic description of nonequilibrium systems comprising many interacting particles. While stochastic thermodynamics has been greatly influential for the study of small systems interacting with their surroundings [8], the thermodynamics of large-scale nonequilibrium systems and their phase transitions have attracted attention only very recently [9][10][11].…”

“…One of the open questions in empirical studies is whether nonequilibrium dynamics and its large entropy production are linked with the critical properties of systems operating near continuous phase transitions [5,6]. Naive mean-field approximations of simple nonequilibrium systems support this assumption.…”

The arrow-of-time in neuroimaging time series identifies causal triggers of brain function

“…Recent neuroimaging endeavours further substantiate this potential: after training a deep learning network to distinguish between temporal segments of forward and backward fMRI time series, Deco et al31 not only observed a variable AoT strength (inferred from classification accuracy on unseen data) across cognitive states, but also between healthy subjects and patients suffering from bipolar disorder, attention deficit hyperactivity disorder or schizophrenia. In another study leveraging the same framework on electrocorticography data, de la Fuente et al44 also revealed that deep sleep and ketamine-induced anesthesia lowered the differences between forward time series and their inverted counterparts, i.e., decreased AoT strength.…”