ConceptNeuro: Neural & Computational Concepts

• ConceptNeuro is a research domain that links semantic concepts to measurable neural substrates and computational models using both biological and artificial systems.
• It integrates findings from single-cell selectivity, distributed neural activations, and cutting-edge neuroimaging to map abstract and concrete concepts.
• Formal frameworks such as conceptors and concept-value networks enable precise, intervention-based decoding and manipulation of concept-driven neural behavior.

Searching arXiv for the cited papers and related concept-neuron work to ground the article in current arXiv records. arxiv_search(query="(Greer, 2019) OR (Ji et al., 26 Mar 2026) OR (Jaeger, 2014) OR (Tapia et al., 2019) OR (Stammer et al., 2024) OR (Hoang-Xuan et al., 2024) OR (Bao et al., 4 Mar 2025) OR (Wang et al., 2 Oct 2025) OR (Kavuri et al., 21 Aug 2025) OR (Liu et al., 2023) OR (Tinauer et al., 2024) OR (Ferrante et al., 2022) OR (Mason et al., 2021) OR (Jaeger, 2014)", max_results=10) Looking up a compact set of the most central papers on concept neurons, concept cells, and conceptor-based neural models. arxiv_search(query="concept neurons concept cells conceptors brain model", max_results=10) CONCEPTNEURO denotes a family of research programs that relate concepts to neural substrates, neural dynamics, or explicit internal variables in artificial systems. Across this literature, concepts are variously modeled as single-neuron selectivities, co-activated neuron groups, recurrent-state filters, object-factor symbols, or statistically localizable activation patterns in fMRI. The unifying theme is the attempt to connect semantic structure to identifiable neural or neural-like mechanisms, whether in biological brains, recurrent reservoirs, diffusion models, LLMs, or concept-guided neuroimaging pipelines (Greer, 2019, Tapia et al., 2019, Jaeger, 2014).

1. Biological grounding: concept cells, conceptual hubs, and high-dimensional selectivity

One major CONCEPTNEURO strand concerns biological concept representation at the single-cell and regional levels. In this line, concept cells—also called “grandmother” or “Jennifer Aniston” cells—are defined as single neurons that fire selectively and invariantly to exemplars of a high-level abstract concept while remaining silent to other stimuli. The theoretical justification offered in “Universal principles justify the existence of concept cells” is that such cells become highly likely in high-dimensional input spaces when neurons have hundreds to thousands of inputs, random initial receptive fields, and local Hebbian-type learning governed by Oja’s rule (Tapia et al., 2019).

That framework makes several explicit quantitative claims. Before learning, at most approximately $1/e \approx 37\%$ of neurons can be strictly single-selective, irrespective of dimension. After learning-induced weight alignment, the probability of silence to random non-target stimuli approaches $1$ as input dimension grows, yielding a selectivity probability $S(n,L)=P^{L-1}$. The paper further states that learning can produce near-perfect selectivity once the fan-in dimension exceeds a modest critical value, with a step-like transition around $n \approx 30$, and that memory capacity grows exponentially in $n$, exceeding $10^{10}$ by $n \approx 60$ in the reported analysis (Tapia et al., 2019). This directly opposes the view that only broad distributed population codes can sustain high-level abstraction.

Regional evidence complements the single-cell account. “Functional Imaging of Conceptual Representations” tested whether conceptual representation requires a multimodal binding region and reported that repetition-suppression effects in perirhinal cortex significantly correlate with behavioral priming across perceptually distinct repetitions. The reported correlation in perirhinal cortex was $r = 0.83$ with $p < 0.0004$, whereas inferior frontal gyrus and superior temporal sulcus did not show significant priming correlations (Awipi, 2012). In this formulation, conceptual representation is not merely feature co-occurrence across distributed sensory cortices; it is associated with a temporal-lobe hub that generalizes across modality.

A further extension of biological CONCEPTNEURO appears in the study of abstract scientific knowledge. “The Neuroscience of Advanced Scientific Concepts” examined 45 physics concepts in 10 Carnegie Mellon physics faculty members and reported that neural representations are organized along dimensions including measurable magnitude, mathematical formulation, periodicity/wave-related structure, and a classical-versus-post-classical dimension associated with reasoning about intangibles, consilience, causal relations not directly observable, and knowledge management of multi-level conceptual structure (Mason et al., 2021). This suggests that concept coding is not confined to concrete object categories; it can also organize highly abstract expert knowledge.

2. Formal frameworks: concept-value networks, conceptors, and compositional mappings

A second strand formalizes how concepts could arise from neural organization. In K. Greer’s “A Concept-Value Network as a Brain Model,” a feature is identified with the physical neural wiring itself, including axon–dendrite connections, while concepts are co-activated groups of neurons sharing a common pattern of features. Features are described as static and “horizontal,” whereas concepts extend “vertically” through the network by aggregating lower-level patterns into higher-level ones (Greer, 2019). The summary accompanying the paper supplies a minimal formalization consistent with that narrative: a joint distribution

$P(F,I)=P(I\mid F)\cdot P(F),$

a distance-dependent synchronization kernel

$1$0

and a concept activation model

$1$1

The same summary notes that glial structures and “signal breaks” are not part of the published paper itself but appear only as natural extensions (Greer, 2019). That distinction is central: the statistical and distance-mediated interpretation is grounded in the paper, whereas glial threshold modulation and compartment boundaries are extrapolations.

Conceptor theory offers a more algebraically explicit mechanism for concept control in recurrent systems. In “Conceptors: an easy introduction,” a conceptor is the regularized linear operator

$1$2

where $1$3 is the state covariance of a reservoir driven by a pattern, and $1$4 is the aperture (Jaeger, 2014). Inserted into the recurrent loop, the conceptor filters state updates so that the network remains inside the state-cloud ellipsoid associated with the target pattern. The theory supports rapid switching among stored patterns, linear morphing between patterns, Boolean-like operations such as OR, AND, and NOT, aperture adaptation, incremental “life-long learning,” and content-addressable recall (Jaeger, 2014). “Controlling Recurrent Neural Networks by Conceptors” develops the same framework further, including autoconceptors, hierarchical denoising, and random-feature conceptors with local scalar updates intended as a more biologically motivated variant (Jaeger, 2014).

A third formal line is provided by “Radically Compositional Cognitive Concepts,” which places concepts in applied category theory. There, concepts are represented in conceptual spaces and composed via functors from syntax to semantics; probabilistic narrative updates are expressed as Markov kernels in $1$5, and a functor

$1$6

translates these kernels into predictive-coding circuits (Smithe, 2019). This shifts CONCEPTNEURO from unit selectivity to compositional structure preservation across representational levels. A plausible implication is that “concept” need not denote a single representational primitive; it can also denote a family of structure-preserving correspondences between semantic composition and neural dynamics.

3. Neuron-level concept interpretation and intervention in artificial networks

In deep-network interpretability, CONCEPTNEURO often refers to methods that assign, verify, or intervene on neuron functions. “Select, Hypothesize and Verify: Towards Verified Neuron Concept Interpretation” explicitly criticizes two assumptions common in prior neuron-description methods: that every neuron has a well-defined discriminative function, and that concepts inferred from high-activation samples are necessarily correct. Its SIEVE framework selects neurons by activation-distribution analysis using

$1$7

forms concept hypotheses from clustered high-activation patches, and then verifies them by generating concept-conditioned images and measuring Activation Rate relative to a top-1% threshold (Ji et al., 26 Mar 2026). Reported mean Activation Rate improvements were approximately $1$8 on ResNet-50/ImageNet, $1$9 on ViT-B/16/ImageNet, and $S(n,L)=P^{L-1}$0 on ResNet-18/Places365, corresponding to roughly $S(n,L)=P^{L-1}$1–$S(n,L)=P^{L-1}$2 gains (Ji et al., 26 Mar 2026).

“LLM-assisted Concept Discovery: Automatically Identifying and Explaining Neuron Functions” removes the fixed concept vocabulary entirely. It defines an ideal concept $S(n,L)=P^{L-1}$3 as one maximizing the probability that the neuron fires more strongly on examples than counterexamples, extracts the strongest exemplars, filters them to a coherent subset of size $S(n,L)=P^{L-1}$4 in CLIP space, asks GPT-4V for a short shared description, and validates the result with synthetic examples and hard co-hyponym counterexamples (Hoang-Xuan et al., 2024). The paper reports coherent textual concepts for 99 of the first 100 final-layer neurons in ResNet50 on ImageNet, compared with 11 for FALCON, and concept-faithfulness scores with median greater than $S(n,L)=P^{L-1}$5 on CLIP-ResNet50 (Hoang-Xuan et al., 2024). The methodological point is that neuron concepts can be open-ended rather than restricted to a predefined ontology.

In LLMs, “NEAT: Concept driven Neuron Attribution in LLMs” defines concept neurons as those whose deactivation most disrupts token probabilities associated with a concept vector built by averaging final-layer hidden states across examples. The principal computational claim is a reduction in required forward passes from $S(n,L)=P^{L-1}$6 to $S(n,L)=P^{L-1}$7, where $S(n,L)=P^{L-1}$8 is the number of neurons and $S(n,L)=P^{L-1}$9 the number of examples (Kavuri et al., 21 Aug 2025). The paper then applies targeted ablations to hate-speech and gender-bias settings, and also reports an Indian-context evaluation in which ablating one male-stereotype neuron and one female-stereotype neuron yields preference percentages of $n \approx 30$0 versus $n \approx 30$1, described as within $n \approx 30$2 percentage points of perfect parity (Kavuri et al., 21 Aug 2025).

Diffusion models add an intervention-oriented variant. “Cones: Concept Neurons in Diffusion Models for Customized Generation” identifies a sparse subset of key/value cross-attention parameters as concept neurons for a subject and represents them with a binary mask over parameters (Liu et al., 2023). Zeroing the corresponding entries is reported to preserve subject generation in new contexts, union masks enable multi-concept generation, and a few further fine-tuning steps support up to four different subjects in one image (Liu et al., 2023). Storage is a central result: Cones stores sparse index sets rather than dense parameter tensors, with reported storage costs from $n \approx 30$3 MB to $n \approx 30$4 MB, versus $n \approx 30$5 MB for Custom Diffusion and $n \approx 30$6 GB for DreamBooth (Liu et al., 2023).

4. Object-factor binding and discrete concept symbols

CONCEPTNEURO is not restricted to single neurons. “Neural Concept Binder” addresses unsupervised object-based visual reasoning by constructing discrete and continuous concept representations called “concept-slot encodings” (Stammer et al., 2024). The soft-binding stage uses SysBinder to produce a block-slot encoding

$n \approx 30$7

where slots correspond to objects and blocks to factors. The hard-binding stage clusters block features from a single-object corpus, distills cluster prototypes into a retrieval corpus, and maps each slot-block vector to a discrete symbol by nearest-neighbor retrieval, yielding

$n \approx 30$8

The framework is explicitly inspectable. The paper describes implicit, comparative, interventional, and similarity inspection modes; symbolic revision can merge, delete, or add cluster exemplars and prototypes; and external knowledge can be introduced by a human or by GPT-4 (Stammer et al., 2024). This makes the learned concept inventory editable rather than merely observable.

The reported benchmark is CLEVR-Sudoku. For $n \approx 30$9 filled cells and five example images per digit, ground-truth concepts solve 100% of puzzles, supervised slot attention reaches 42% on CLEVR-Easy and 74% on full CLEVR, SysBinder discretized by $n$0 solves 0%, and unsupervised NCB solves 55% and 27%, respectively. After human revision of the retrieval corpus, NCB improves to 67% and 45% (Stammer et al., 2024). The same paper reports a digit-classification error reduction from approximately 88% for the naive SysBinder discretization to approximately 2% for NCB, and a qualitative example in which merging two purple clusters raises digit accuracy from 94% to 98% and solved rate from 55% to 67% on CLEVR-Easy (Stammer et al., 2024). In this formulation, concept binding is a discrete symbolic layer built on top of continuous latent factors.

5. Concept localization and decoding in neuroimaging

A neuroimaging-oriented CONCEPTNEURO literature seeks to localize or reconstruct concept-selective activity patterns from brain data. The strongest early evidence for multimodal conceptual localization comes from the perirhinal cortex study already noted, where cross-modal repetition suppression tracked conceptual priming rather than perceptual repetition alone (Awipi, 2012). A related but more abstract line appears in the physics-concept study, where Gaussian naïve Bayes classification of activation patterns achieved mean normalized rank accuracy of 0.78 within participant and 0.70 across participants, and regression from semantic ratings predicted held-out activation patterns with mean $n$1 (Mason et al., 2021). Together, these works treat concept structure as decodable from distributed but stable fMRI signatures.

“Semantic Brain Decoding: from fMRI to conceptually similar image reconstruction of visual stimuli” implements a decoding pipeline from visual-cortex fMRI to the last convolutional layer of ResNet50, followed by nearest-neighbor semantic labeling and conditioning of a latent diffusion model (Ferrante et al., 2022). On the GOD dataset, the reported average Wu–Palmer similarity on the test set is $n$2, and a human evaluation identified the model reconstructions as the better match on 81% $n$3 4% of test trials (Ferrante et al., 2022). This is concept-level reconstruction rather than pixel-wise inversion: the generated images are intended to preserve semantic and contextual similarity.

“MindSimulator: Exploring Brain Concept Localization via Synthetic FMRI” moves from decoding to generative hypothesis formation. It models the conditional distribution of fMRI given visual stimulus with a latent-diffusion system aligned to CLIP image tokens and uses synthetic recordings to run voxelwise one-sample $n$4-tests for concept localization (Bao et al., 4 Mar 2025). Reported synthesis fidelity includes voxel-level Pearson-$n$5 up to 0.355 and semantic reconstruction metrics such as PixCorr = 0.201 and SSIM = 0.298, while region-localization examples on subject 01 show places accuracy improving from 32.0% for a linear baseline to 52.6% for MindSimulator, and bodies from 45.8% to 91.7% (Bao et al., 4 Mar 2025). The paper also reports predicted surfer-, plane-, food-, and bed-selective patches in higher visual cortex (Bao et al., 4 Mar 2025).

A disease-classification analogue appears in “Explainable concept mappings of MRI,” which trains a relevance-regularized 3D CNN on quantitative $n$6 maps to distinguish Alzheimer’s disease from normal controls and then extracts concept-conditional relevance maps with Concept Relevance Propagation (Tinauer et al., 2024). The selected model is reported at balanced accuracy $n$7, sensitivity $n$8, specificity $n$9, and AUC $10^{10}$0, and the most relevant concepts are reported primarily in and adjacent to the basal ganglia, consistent with conventional ROI analysis and established histological findings (Tinauer et al., 2024). Here, “concept” refers to hidden-channel relevance patterns rather than linguistically named categories.

6. Clinical translation, misconceptions, and open problems

The most explicitly clinical use of the term appears in “Interpretable Neuropsychiatric Diagnosis via Concept-Guided Graph Neural Networks,” which names its framework CONCEPTNEURO and defines concepts as structured functional-connectivity subgraphs linking region sets with a hyper- or hypo-connectivity direction prior (Wang et al., 2 Oct 2025). Candidate phrases are generated by prompting GPT-4.1 with disorder-specific NeuroQuery terms, parsed into atlas-aligned ROI sets, embedded with a small GNN, and passed through a sparse, sign-constrained concept bottleneck before classification. On ABCD under a GCN backbone, the reported binary accuracies improve from 61.2 ± 1.7 to 65.8 ± 1.0 for Anxiety, 70.1 ± 1.1 to 76.1 ± 1.7 for OCD, 57.8 ± 1.8 to 62.6 ± 1.3 for ADHD, 65.5 ± 1.3 to 67.5 ± 1.9 for ODD, and 69.3 ± 1.5 to 72.3 ± 1.2 for Conduct; analogous gains are reported on HCP-D and with GAT, GraphSAGE, and GIN backbones (Wang et al., 2 Oct 2025). Expert validation reports up to 80% top-10 concept agreement and more than 75% per-subject top-10 ranking agreement (Wang et al., 2 Oct 2025).

Several recurrent misconceptions are corrected by this literature. First, not every neuron appears to have a well-defined, discriminative concept; SIEVE explicitly argues that redundant or weakly selective neurons can yield spurious interpretations, so observation without intervention is inadequate (Ji et al., 26 Mar 2026). Second, concept representation is not reducible to a single format. Some papers emphasize single neurons or tightly selective cells (Tapia et al., 2019), whereas Greer’s framework states that features can be distributed entities and not concentrated to a single area (Greer, 2019). Third, fixed vocabularies are not compulsory. Open-ended LLM-assisted discovery is motivated precisely by the limitation of predefined concept sets and curated exemplars (Hoang-Xuan et al., 2024).

The principal unresolved issues are methodological rather than terminological. Verification pipelines depend on generative models whose training distributions may bias example quality under domain shift, as noted explicitly for SIEVE and for synthetic brain-recording generation (Ji et al., 26 Mar 2026, Bao et al., 4 Mar 2025). Fixed concept pools can miss specialized or unseen concepts, while open-ended MLLM pipelines can refuse answers, hallucinate abstractions, or incur substantial cost (Ji et al., 26 Mar 2026, Hoang-Xuan et al., 2024). In biological modeling, some proposed mechanisms—such as glial modulation of thresholds or explicit “signal breaks” for compartmentalization—are only natural extensions and not part of the original published model (Greer, 2019). This suggests that CONCEPTNEURO is best understood not as a single theory, but as a research domain centered on one question: how concepts can be made operationally identifiable within neural organization, neural dynamics, or neural-network computation.