arXiv subjects

Christian Meisel

Publications and source records attributed to Christian Meisel.

9 recordsLinked to original sources

arXiv2026

Adaptive self-organized criticality in deep neural networks

Deep neural networks are high-dimensional dynamical systems whose function depends on the stable propagation of activity and perturbations across many layers. Maintaining suitable dynamical regimes may therefore be essential for robust learning and for preventing dynamical instabilities during training. Here, we show that the global dynamical state of a deep neural network can be autonomously regulated by purely local homeostatic plasticity. Neuronal activity is inferred from responses across inputs, and individual synapses are strengthened or weakened using only the activity of their postsynaptic neuron. Without measuring any global network property, this rule drives networks from both subcritical and supercritical initial conditions toward a common critical state, characterized by conserved activity propagation and a vanishing largest finite-time Lyapunov exponent. When combined with gradient-based learning, homeostatic adaptation counteracts the training-induced drift toward supercritical dynamics, while revealing a competition between dynamical regulation and task optimization. Our results demonstrate how adaptive self-organization can be implemented in deep neural networks and how local plasticity can control their collective dynamical operating point.

q-bio.NC↗

arXiv2025

Critical dynamics governs deep learning

The rapid advances in artificial intelligence (AI) have largely been driven by scaling deep neural networks (DNNs) - increasing model size, data, and computational resources. Yet performance is ultimately governed by network dynamics. The lack of a principled understanding of DNN dynamics beyond heuristic design has contributed to challenges in robustness, suboptimal performance, high energy consumption, and pathologies in continual and AI-generated content learning. In contrast, the human brain appears largely resilient to these problems, and converging evidence suggests this advantage arises from dynamics poised at a critical phase transition. Inspired by this principle, we propose that criticality provides a unifying framework linking structure, dynamics, and function in DNNs. First, analyzing more than 80 state-of-the-art models, we show that a decade of AI progress has implicitly driven successful networks toward criticality - explaining why some architectures succeeded while others failed. Second, we demonstrate that explicitly incorporating criticality into training improves robustness and accuracy while mitigating key limitations of current models. Third, we show that major AI pathologies - including performance degradation in continual learning and model collapse during training on AI-generated data - reflect a loss of critical dynamics. By maintaining networks near criticality, we provide a principled solution to these failures, demonstrating that criticality-based optimization prevents degradation and collapse. Our results establish criticality as a substrate-independent principle of intelligence, connecting AI progress with fundamental principles of brain function, and offering both theoretical insight and practical strategies to ensure long-term DNN performance and resilience as models scale.

q-bio.NC↗

arXiv2019

Deep learning from wristband sensor data: towards wearable, non-invasive seizure forecasting

Seizure forecasting may provide patients with timely warnings to adapt their daily activities and help clinicians deliver more objective, personalized treatments. While recent work has convincingly demonstrated that seizure risk assessment is possible, these early approaches relied largely on complex, often invasive setups including intracranial electrocorticography, implanted devices and multi-channel EEG, which limits translation of these methods to broad clinical application. To facilitate broader adaptation of seizure forecasting in clinical practice, non-invasive, easily applicable techniques that reliably assess seizure risk, in combination with clinical information, are crucial. Wristbands that continuously record physiological parameters, including electrodermal activity, body temperature, blood volume pressure and actigraphy, may afford monitoring of autonomous nervous system function and movement relevant for such a task, hence minimizing potential complications associated with invasive monitoring, and avoiding stigma associated with bulky external monitoring devices on the head. Here, we use deep learning to analyze long-term, multi-modal wristband sensor data from 50 patients with epilepsy (total duration $>$1400 hours) to assess its capability to distinguish preictal from interictal states. Prediction performance is assessed using area under the receiver operating charateristic (AUC) and improvement over chance (IoC) based on F1 scores. Using one- and two-dimensional convolutional neural networks, we identified better-than-chance predictability in out-of-sample test data in 60\% of the patients in leave-one-out and 43\% of patients in pseudo-prospective approaches. These results provide a step towards developing easier to apply, non-invasive methods for seizure risk assessments in patients with epilepsy.

q-bio.NC↗

arXiv2019

Antiepileptic drugs induce subcritical dynamics in human cortical networks

Cortical network functioning critically depends on finely tuned interactions to afford neuronal activity propagation over long distances while avoiding runaway excitation. This importance is highlighted by the pathological consequences and impaired performance resulting from aberrant network excitability in psychiatric and neurological diseases, such as epilepsy. Theory and experiment suggest that the control of activity propagation by network interactions can be adequately described by a branching process. This hypothesis is partially supported by strong evidence for balanced spatiotemporal dynamics observed in the cerebral cortex, however, evidence of a causal relationship between network interactions and cortex activity, as predicted by a branching process, is missing in humans. Here we test this cause-effect relationship by monitoring cortex activity under systematic pharmacological reduction of cortical network interactions with antiepileptic drugs. We report that cortical activity cascades, presented by the propagating patterns of epileptic spikes, as well as temporal correlations decline precisely as predicted for a branching process. Our results provide the missing link to the branching process theory of cortical network function with implications for understanding the foundations of cortical excitability and its monitoring in conditions like epilepsy.

q-bio.NC↗

arXiv2017

Decline of long-range temporal correlations in the human brain during sustained wakefulness

Sleep is crucial for daytime functioning, cognitive performance and general well-being. These aspects of daily life are known to be impaired after extended wake, yet, the underlying neuronal correlates have been difficult to identify. Accumulating evidence suggests that normal functioning of the brain is characterized by long-range temporal correlations (LRTCs) in cortex, which are supportive for decision-making and working memory tasks. Here we assess LRTCs in resting state human EEG data during a 40-hour sleep deprivation experiment by evaluating the decay in autocorrelation and the scaling exponent of the detrended fluctuation analysis from EEG amplitude fluctuations. We find with both measures that LRTCs decline as sleep deprivation progresses. This decline becomes evident when taking changes in signal power into appropriate consideration. Our results demonstrate the importance of sleep to maintain LRTCs in the human brain. In complex networks, LRTCs naturally emerge in the vicinity of a critical state. The observation of declining LRTCs during wake thus provides additional support for our hypothesis that sleep reorganizes cortical networks towards critical dynamics for optimal functioning.

q-bio.NC↗

arXiv2017

The interplay between long- and short-range temporal correlations shapes cortex dynamics across vigilance states

Increasing evidence suggests that cortical dynamics during wake exhibits long-range temporal correlations suitable to integrate inputs over extended periods of time to increase the signal-to-noise ratio in decision-making and working memory tasks. Accordingly, sleep has been suggested as a state characterized by a breakdown of long-range correlations; detailed measurements of neuronal timescales that support this view, however, have so far been lacking. Here we show that the long timescales measured at the individual neuron level in freely-behaving rats during the awake state are abrogated during non-REM (NREM) sleep. We provide evidence for the existence of two distinct states in terms of timescale dynamics in cortex: one which is characterized by long timescales which dominate during wake and REM sleep, and a second one characterized by the absence of long-range temporal correlations which characterizes NREM sleep. We observe that both timescale regimes can co-exist and, in combination, lead to an apparent gradual decline of long timescales during extended wake which is restored after sleep. Our results provide a missing link between the observed long timescales in individual neuron fluctuations during wake and the reported absence of long-term correlations during deep sleep in EEG and fMRI studies. They furthermore suggest a network-level function of sleep, to reorganize cortical networks towards states governed by slow cortex dynamics to ensure optimal function for the time awake.

q-bio.NC↗

arXiv2011

Failure of adaptive self-organized criticality during epileptic seizure attacks

Critical dynamics are assumed to be an attractive mode for normal brain functioning as information processing and computational capabilities are found to be optimized there. Recent experimental observations of neuronal activity patterns following power-law distributions, a hallmark of systems at a critical state, have led to the hypothesis that human brain dynamics could be poised at a phase transition between ordered and disordered activity. A so far unresolved question concerns the medical significance of critical brain activity and how it relates to pathological conditions. Using data from invasive electroencephalogram recordings from humans we show that during epileptic seizure attacks neuronal activity patterns deviate from the normally observed power-law distribution characterizing critical dynamics. The comparison of these observations to results from a computational model exhibiting self-organized criticality (SOC) based on adaptive networks allows further insights into the underlying dynamics. Together these results suggest that brain dynamics deviates from criticality during seizures caused by the failure of adaptive SOC.

q-bio.NC↗

arXiv2011

On spatial and temporal multilevel dynamics and scaling effects in epileptic seizures

Epileptic seizures are one of the most well-known dysfunctions of the nervous system. During a seizure, a highly synchronized behavior of neural activity is observed that can cause symptoms ranging from mild sensual malfunctions to the complete loss of body control. In this paper, we aim to contribute towards a better understanding of the dynamical systems phenomena that cause seizures. Based on data analysis and modelling, seizure dynamics can be identified to possess multiple spatial scales and on each spatial scale also multiple time scales. At each scale, we reach several novel insights. On the smallest spatial scale we consider single model neurons and investigate early-warning signs of spiking. This introduces the theory of critical transitions to excitable systems. For clusters of neurons (or neuronal regions) we use patient data and find oscillatory behavior and new scaling laws near the seizure onset. These scalings lead to substantiate the conjecture obtained from mean-field models that a Hopf bifurcation could be involved near seizure onset. On the largest spatial scale we introduce a measure based on phase-locking intervals and wavelets into seizure modelling. It is used to resolve synchronization between different regions in the brain and identifies time-shifted scaling laws at different wavelet scales. We also compare our wavelet-based multiscale approach with maximum linear cross-correlation and mean-phase coherence measures.

q-bio.NC↗

arXiv2009

Adaptive self-organization in a realistic neural network model

Information processing in complex systems is often found to be maximally efficient close to critical states associated with phase transitions. It is therefore conceivable that also neural information processing operates close to criticality. This is further supported by the observation of power-law distributions, which are a hallmark of phase transitions. An important open question is how neural networks could remain close to a critical point while undergoing a continual change in the course of development, adaptation, learning, and more. An influential contribution was made by Bornholdt and Rohlf, introducing a generic mechanism of robust self-organized criticality in adaptive networks. Here, we address the question whether this mechanism is relevant for real neural networks. We show in a realistic model that spike-time-dependent synaptic plasticity can self-organize neural networks robustly toward criticality. Our model reproduces several empirical observations and makes testable predictions on the distribution of synaptic strength, relating them to the critical state of the network. These results suggest that the interplay between dynamics and topology may be essential for neural information processing.

q-bio.NC↗