Causal interactions in complex systems like the brain can be examined at different spatiotemporal levels. While it is often assumed that the micro level is causally complete, this work shows that causal power can be stronger at macro levels. Using a measure called ΦMax, developed within integrated information theory, the authors systematically evaluated causal power at micro and macro levels in simplified neuronal-like systems. For systems with indeterminism or degeneracy, ΦMax peaked at a macro level when coarse-graining micro elements produced macro mechanisms with high irreducible causal selectivity.
A formal analysis shows that many contemporary theories of consciousness fail to meet the requirements of falsifiability and non-triviality. This is especially problematic for claims that Large Language Models (LLMs) might be conscious: because functionally equivalent systems exist, no falsifiable and non-trivial theory can judge them conscious, forming a disproof of LLM consciousness. In contrast, theories of consciousness that require continual learning do satisfy these formal constraints for humans. This supports the hypothesis that if continual learning is linked to human consciousness, the current lack of continual learning in LLMs is intimately tied to their lack of consciousness.
Dreams may serve a biological function similar to how noise injections help deep neural networks avoid overfitting. Overfitting occurs when a system performs well on learned data but fails to generalize to new situations. The overfitted brain hypothesis proposes that dreams generate corrupted sensory inputs through stochastic neural activity, helping the brain generalize from daily learning and preventing overfitting. Dream loss, as opposed to sleep loss more broadly, would produce an overfitted brain that can memorize but cannot generalize appropriately. The hypothesis is compared with existing dream theories, and evidence from neuroscience and deep learning is reviewed, along with testable predictions for future research.
Falsification, a cornerstone of scientific testing, is especially problematic for theories of consciousness. In the standard experimental setup, a theory's predicted experience (based on brain data) is compared with an inferred experience (based on report or behavior). If inference and prediction are independent, any minimally informative theory is automatically falsified—a dilemma for many current theories that rely on report to infer conscious experience. If inference and prediction are strictly dependent, the theory becomes unfalsifiable. The paper explores potential ways out of this dilemma, highlighting a fundamental challenge for empirical testing in consciousness research.
Neuroscience lacks a unifying theoretical framework, remaining pre-paradigmatic in the Kuhnian sense despite methodological advances. The essay argues that consciousness should serve as the missing organizing principle, analogous to evolution in biology. Evidence of reproducibility failures in neuroimaging and the field's reliance on 'cool' rather than fundamental findings supports this call to place consciousness at the center of neuroscience.