A new study from Tamas Spisak and Karl Friston reveals that attractor neural networks—a key feature of brain dynamics—can self-organize from the free energy principle, a fundamental theory of brain function. Published on arXiv in May 2025 and updated in May 2026, the work shows that such networks automatically learn to orthogonalize attractor states, maximizing information storage and generalization.
The research
The team applied the free energy principle to a universal partitioning of random dynamical systems. They derived that attractor networks emerge without explicit learning rules, instead following biologically plausible inference and learning dynamics. Mathematically, the networks favor nearly orthogonal attractor representations, which efficiently span the input subspace. This enhances mutual information between hidden causes and sensory data. Simulations confirmed that random data presentation yields symmetric, sparse couplings, while sequential data produces asymmetric couplings and non-equilibrium dynamics—a generalization of Boltzmann Machines.
Why it matters
These findings unify self-organization in neural networks with Bayesian active inference. For readers, it suggests that the brain naturally structures memories to avoid overlap, improving recall and generalization. Understanding this could inspire better AI architectures and clarify how our own brains optimize learning.
What you can do
To support your brain's natural learning, expose yourself to varied environments and sequential tasks that encourage orthogonal memory formation. Consistent, spaced practice helps reinforce distinct neural patterns.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
