¶

Large-scale brain attractor networks¶

We work on a scale-free, first-principles computational model of large-scale brain dynamics. The framework links connectivity to activity and can disentangle network-level computations at a given level of description (imaging modality and parcellation) from intrinsic computations and activity at lower levels — including sensory inputs and recurrent low-level processing.

Approach¶

Starting from the Free Energy Principle applied to deep particular partitions, attractor networks emerge without hand-crafted learning or inference rules. Attractors on the free-energy landscape act as Bayesian priors; inference integrates data into posteriors; learning tunes couplings to reduce long-term surprise. A hallmark is self-orthogonalization: approximately orthogonal attractor representations that efficiently span the input subspace.

Mapped to neuroimaging, resting-state and task fMRI can be fit (e.g. via score matching) onto connectivity-based attractor networks. Resting-state networks appear as approximate attractors; task dynamics can be read as Bayesian inference in which weights capture network-level dynamics and biases capture upstream / mesoscale drive. Clinical states (e.g. acute vs chronic pain) can be framed as changes in landscape geometry — shallow vs deep, trapping wells.

Key publications & resources¶

• Spisak & Friston — Self-orthogonalizing attractor neural networks emerging from the free energy principle
  Neurocomputing · Manuscript site · ScienceDirect

• Englert et al. — Functional Connectivity-based Attractor Dynamics in Rest, Task, and Disease
  eLife · connattractor site · GitHub

Software¶

• connattractor — connectome-based Hopfield / attractor analyses for fMRI (pip install connattractor)

• attractome — connectivity-based FEP attractor nets for parcellated fMRI (HRF deconvolution, score matching, attractor utilities; research preview)

See also Brain models / fcHNN and Publications.