Twelve proposed correspondences between human brain structures and components of artificial neural networks. Anatomy is segmented from published probabilistic atlases in MNI152 space, so position and relative size are measured rather than drawn. Each correspondence holds only if the system being compared has particular properties; change the configuration below and the map changes with it.
Switching a property off removes the thing in the artificial system that the correspondence was about. The anatomy never changes.
Each card is one component of the artificial system and the brain structure it is claimed to correspond to. The colour swatch is that structure's colour in the anatomy view. A faded card has no counterpart in the current configuration. The word at bottom right is the evidence grade, with an arrow if the configuration has weakened it. Click a card to inspect the correspondence.
Why these comparisons are made at all. Marr's three levels of analysis separate what a system computes, the algorithm and representation it uses, and how that is physically implemented (Marr, Vision, 1982). Two systems can match at one level and diverge completely at another, so every row states the level at which its correspondence is claimed. A match at the computational level — both systems solve the same problem — says nothing about shared mechanism.
Evidence grades. Strong means a quantitative comparison exists in which model responses predict measured neural responses, or a formal quantity was predicted and then recorded. Partial means a shared computational principle with a demonstrably different implementation. Loose means the correspondence is conceptual or terminological, with no model-to-brain measurement. With everything switched on, only three of twelve rows are strong, and two of those are in vision.
What the configuration does. Switching a property off does not change the brain. It removes the thing in the artificial system that the correspondence was about, so the row either loses its counterpart entirely or falls to a weaker grade, and the tool states which. Imposing a biological constraint runs the comparison the other way: it asks which artificial mechanisms would have to be abandoned. The dependency conditions and grade transitions are the author's judgement, stated as such, in the same way as the match ratings; the underlying findings are cited per row.
Two cautions a professional reader should apply throughout. Reverse inference — concluding that a cognitive process is engaged because a region activated — is not deductively valid, and is weakest for regions that activate across many tasks, such as anterior cingulate and insula (Poldrack 2006, Trends Cogn Sci). Separately, high brain-prediction scores do not license claims of mechanistic similarity: networks sharing little with biological vision can nonetheless predict neural data well (Bowers et al. 2023, Behav Brain Sci).
Representational alignment is not biological plausibility. Studies that train a network and then test whether its representations predict measured brain signals — encoding and decoding work in vision and language — answer one question: do the two systems represent the same things. They do not answer whether a neuron could implement the mechanism at all, which is a separate literature about weight transport, local learning rules and weight sharing. A model can predict brain data very well while being biologically impossible, and the reverse. This tool keeps the two apart: the evidence grade on each row reports alignment, while the biological constraints in the configuration panel test plausibility.
Anatomy provenance. Surfaces are extracted by marching cubes from the FSL MNI152 1 mm template and from maximum-probability labellings (threshold 25%) of the Harvard-Oxford cortical and subcortical atlas and the Julich-Brain cytoarchitectonic atlas (Amunts, Zilles and colleagues), with cerebellar lobules from the AAL atlas. Structures are shown at true relative size and position in MNI millimetre space. Atlas-derived volumes run systematically larger than manual tracing because probabilistic boundaries are inclusive; published volumetry is given alongside, with its method, so the two are never confused. A template is an average of many brains, and individual anatomy — the language network above all — varies substantially (Fedorenko et al. 2010, J Neurophysiol).