Children learn grammar from exposure to sentences — no explicit rules, no error signal, no backpropagation. The NEMO model (Mitropolsky & Papadimitriou, 2025) demonstrates that the same is possible in the Assembly Calculus: a network of brain areas implementing six neuroscience principles — excitatory neurons, brain areas, random synapses, Hebbian plasticity, local inhibition, and inter-area inhibition — can acquire word semantics, noun/verb classification, and constituent order from grounded sentences, with the same code learning SVO and SOV order depending only on the training data. No language-specific parameters, no grammar file, no optimizer. This page describes the model, the neurolinguistic predictions it generates, and the experimental evidence from both the published papers and our simulation platform.
The NEMO architecture
NEMO (NEural MOdel of language) extends the Assembly Calculus with a multi-area architecture designed around the functional requirements of language processing. The 2023 version defined the architecture; the 2025 version demonstrated that grammar can be learned, not hardcoded:
| Area | Function |
|---|---|
| Phon | Phonological input — pre-initialized assemblies for each word |
| Lex1, Lex2 | Lexical areas for nouns and verbs (learned via Hebbian plasticity) |
| Visual, Motor | Grounding: concrete noun referents (Visual), action verb referents (Motor) |
| C1…Cm | Context areas (olfactory, emotional, locational features) |
| RoleAgent, RoleAction, RolePatient | Thematic roles — who does what to whom |
| Subj, Verb, Obj | Syntactic constituent areas — learned constituent order |
| Mood | Grammatical mood (declarative, interrogative) |
The key innovation: grammar emerges from Hebbian learning over grounded sentences. Training proceeds in two phases. In the first phase (word learning), the network is exposed to two-word sentences with simultaneous grounding — the phonological form of each word alongside its visual or motor referent, forming bidirectional Hebbian pathways between Phon, Lex1/Lex2, and the sensory areas. Approximately 10 sentences per word are needed for stable assembly formation, with the total growing linearly with lexicon size. In the second phase (syntax learning), the network encounters full sentences with thematic roles, and the Hebbian pathways between lexical areas, role areas, and syntactic constituent areas strengthen to encode constituent order. The same architecture trained on SVO data learns SVO; trained on SOV data, it learns SOV. The grammar is in the data, not the code.
Developmental curriculum
The 2025 paper defines two training phases: word learning (single words with grounding) and syntax learning (full sentences with thematic roles). Our simulation platform extends this with a four-stage developmental curriculum modeled on child language acquisition, where matching the developmental trajectory is itself a test of the model's biological plausibility:
| Stage | Age equivalent | Vocabulary | Structures |
|---|---|---|---|
| 1 | 12–18 months | ~50 words | Single-word naming (maps to paper's word learning phase) |
| 2 | 18–24 months | ~300 words | Vocabulary spurt, two-word combos |
| 3 | 24–30 months | ~500 words | Telegraphic speech, SVO emerging |
| 4 | 30–36 months | ~1,000 words | Full sentences (maps to paper's syntax learning phase) |
The paper reports that word learning requires approximately 10 training sentences per word, with the total growing linearly with lexicon size. Constituent order learning also scales linearly. In our extended curriculum, word order is correctly learned by Stage 3 for SVO training data. The grounding system links words to sensory features — the assembly for a noun like “dog” shares neurons with both the visual referent assembly (activated by seeing the object) and the phonological assembly (activated by hearing the word), so that perception and language activate overlapping neural populations.
N400 = global pre-k-WTA energy
The following is from our simulation experiments on this platform, not from the published NEMO papers. It is a prediction of the Assembly Calculus that we have validated computationally but not yet published.
The N400 is a well-established event-related potential (ERP) component: a negative deflection peaking around 400ms after stimulus onset, larger for semantically unexpected words. It is one of the most replicated findings in cognitive neuroscience. The question has always been: what computation does it reflect?
Our finding: the N400 maps to global pre-k-WTA energy — the total synaptic input summed across all neurons in a brain area before winner-take-all selection. The mechanism is input redundancy: when a prime (“nurse”) is semantically related to the target (“doctor”), their assemblies share neurons. The shared neurons receive input from both the stimulus and the recurrent connections, but synaptic weights saturate — overlapping inputs are capped rather than summing freely. Total energy is lower. When the prime is unrelated (“table” before “doctor”), the prime and target activate disjoint neuron sets; inputs add without saturation, producing higher total energy and a larger N400.
The effect is robust across parameter sweeps, vocabulary sizes, and compute engines. Because the model is a deterministic simulation rather than noisy human EEG, effect sizes are very large (Cohen's d ≈ −25) — this reflects the signal-to-noise advantage of simulation, not a claim about human cortical effect magnitudes. For the full derivation, robustness analysis, and connection to the N400 literature, see the dedicated N400 page.
P600 and the triple dissociation
Also from our simulation experiments, not yet published.
The P600 is an ERP component with onset around 500ms and peak around 600ms, associated with syntactic processing — particularly structural violations and reanalysis. In the Assembly Calculus, it maps to assembly instability: the Jaccard distance between successive winner sets across settling rounds in structural areas (ROLE_AGENT, ROLE_PATIENT, VP). When structural expectations are violated — for instance, a noun appearing where the network expects a verb — the structural assemblies rather than converging. When pathways have been consolidated through Hebbian training, the assemblies settle quickly; when the input violates learned structure, they do not.
The strongest result is a triple dissociation — three independent metrics from a single architecture, each selectively sensitive to a different type of violation, mapping onto Friederici's (2002) three-phase language processing model: Phase 1, initial structure building (~100–300ms, corresponding to the ELAN), maps to core-area instability; Phase 2, lexical-semantic integration (~300–500ms, the N400), maps to global pre-k-WTA energy; Phase 3, syntactic reanalysis (~500–1000ms, the P600), maps to structural-area . Semantic violations elevate N400 energy but leave structural instability unchanged; category violations elevate structural instability but leave N400 energy unchanged. No separate “semantic module” or “syntax module” is needed — the dissociation emerges from the same Hebbian dynamics operating on different brain areas at different timescales. For the full data, per-area breakdown, and morphosyntactic agreement results, see the ERP triple dissociation page.
Cross-linguistic and open questions
The 2023 paper demonstrated rule-based parsing with pre-existing word representations. The 2025 paper demonstrates learned grammar across word orders, testing SVO (the dominant order in English, French, Chinese, Swahili) and SOV orders, and noting that OSV and OVS orders — which are extremely rare in natural languages — require significantly more training data. This typological prediction (common orders are easier to learn) is a natural consequence of the architecture, not a parameter choice.
The deeper question: can statistical learning over grounded sentences explain cross-linguistic universals without innate Universal Grammar? If the Assembly Calculus naturally converges on certain grammatical structures (hierarchical composition, argument structure, agreement) from distributional statistics alone, that is evidence for the sufficiency of domain-general learning mechanisms. If it cannot — if certain structures require architectural priors — that is evidence for the necessity of language-specific neural circuitry. Either answer is informative.