The Implicit Brain Model is a multimodal foundation model of whole-brain state and dynamics. One parameter set is declared once, and task-specific models are materialized out of it on demand. Because the representation is tied to the brain rather than to an instrument, film, MEG, EEG and motion capture all constrain the same parameters — data that could not otherwise be combined at all.
Specify a specific region, resolution, or bandwidth and only the processes and variables that actually matter materialize. The same brain can be coarse everywhere and fine around a single electrode.
M = materialize(R, r, B, F, A, T, P)
A brain's rhythms are not decoration laid over its anatomy — they are what a loop with a delay and a gain does. IBM-1 declares the loops themselves: which structures, in which order, with what conduction delay. A band is then a consequence to be checked, not a spectrum to be fitted.
The loop the whole brain integrates through: cortex drives thalamus, and thalamus drives cortex back.
Motor and premotor cortex propose; the striatum and pallidum release one proposal and suppress the rest; the thalamus returns the winner. The goal is that action selection falls out of a loop that was declared, rather than out of a policy that was trained.
Entorhinal input rides a slow theta cycle, and each cycle is divided into faster gamma slots. The goal is that the order of an episode is carried by when something fires inside the cycle, rather than stored beside it as an index.
Occipito-temporal cortex proposes what is worth looking at, parietal cortex holds the map it is proposed on, and frontal cortex commits to one place. The goal is that visual salience is a competition running inside the substrate, not a saliency map bolted on.
The amygdala scores what has just arrived, ventromedial prefrontal cortex and the anterior cingulate revise that score against context, and the accumbens turns it into a pull toward or away. The goal is that value is a state of a brain attached to a body, not a scalar handed in from outside it.
Motor cortex sends a copy of each command through the pons to the cerebellum, which returns a correction through the thalamus before the movement has finished. The goal is timing and error correction that come from the loop, not from a tuned controller.
Posterior cingulate, precuneus, medial prefrontal and angular cortex rise together when nothing outside demands attention. The goal is a resting brain with structure — recollection and self-reference — rather than a model that idles between inputs.
Cochlea, brainstem, medial geniculate and Heschl's gyrus, each station locking to the timing of the sound. The goal is that speech is followed by a loop that entrains to it, not by a spectrogram handed in from outside.
The ascending arousal system sets how the thalamus gates the cortex, and so whether the brain is awake, drowsy or asleep. The goal is that sleep and waking are states the substrate moves between, not modes chosen by a flag.
Retina to the lateral geniculate to striate cortex and out into the visual hierarchy, with the thalamus gating what gets through. The goal is that eyes open and eyes closed are one loop in two states.
Sensorimotor cortex idles in a rhythm that breaks the moment a movement is planned, made or watched. The goal is that preparing to act shows in the loop before any muscle moves.
Each figure lights the structures the substrate actually declares for that loop and runs a wave around it. The on-screen rate is a depiction — every band slowed by the same factor so that none of them alias — and is not a measured frequency.
Every system the atlas carries — skeleton, muscle, vessels, viscera and the skin over them — on one driven skeleton, with the cortex closing the loop out to the muscle and back.
Control: scramble what the cortex learned and it still stands. The cortex owns the muscles; that is not the same as it knowing how to walk.
!pip install git+https://github.com/JacobFV/IBM-1import ibm
from ibm.materialize import library
ibm.load_all(seal=True, strict=True)
print(ibm.REGISTRY.summary())13 fields · 113 components · 22 supports · 20 anatomical systems 17 topologies · 30 processes · 115 implementations
m = library.MODELS["eeg_forward"]
print(m.request.describe())request 'eeg-forward' subject=template frame=subject_t1 targets device.contact_potential, electromagnetic.potential regions cortex, conductor r(q) 3 rules, default 6 mm B(q) default Band(0.5-100Hz), 3 overrides window n=2048 dt=0.001s (2.048s, nyquist 500 Hz)
# one kernel carries the parameters for any materialization
from huggingface_hub import hf_hub_download
ckpt = hf_hub_download("jacob-valdez/ibm-1",
"implicit/ibm1.implicit.s30k.e128.fused34.pt")