Cognitive and Neural Evidence

• Cognitive and neural evidence is the integration of behavioral, computational, and neurophysiological data to reveal how mental processes arise from underlying brain activity.
• Parametric models such as the drift-diffusion model, combined with EEG and fMRI analyses, quantitatively map latent cognitive variables to neural signals.
• Generative network models and advanced decoding techniques provide mechanistic insights linking connectivity, long-range wiring, and neural representations to cognitive performance.

Cognitive and neural evidence encompasses the integration of behavioral, computational, and neurophysiological data to elucidate how mental processes arise from brain activity. This domain unites mechanistic models, neural measurements, and cognitive task analyses to bridge theoretical constructs—such as evidence accumulation, working memory capacity, network efficiency, representational geometry, and decision strategies—with direct observations from brain signals. Approaches span biophysical modeling, network generativity, neural encoding/decoding, and quantitative behavioral-neural alignments, yielding rigorous indices of cognitive mechanisms as instantiated in neural circuits.

1. Parametric Models Linking Behavior to Neural Dynamics

Quantitative frameworks such as the drift–diffusion model (DDM) provide a principled mapping from choice-behavior (response times, accuracy) to latent cognitive parameters (drift rate $\delta$, boundary separation $\alpha$, starting point $z$) with explicit neural correlates. The Wiener first-passage time (WFPT):

$$p\bigl(t\mid \delta,\alpha,z\bigr) = \frac{\pi}{\alpha^2} \exp\!\left(-\delta \alpha z - \frac{\delta^2}{2}(t-\tau)\right) \sum_{m=-M}^M (z+2m) \exp\!\left(-\frac{(z+2m)^2}{2(t-\tau)}\right)$$

links RT to cognitive latent variables.

Recent advances directly predict DDM parameters from single-trial EEG using shallow neural architectures (Decision SincNet). The model leverages Sinc-convolution layers to learn band-pass filters that extract relevant spectral components; subsequent spatial convolution, temporal pooling, and specialized readouts yield per-trial $\hat\delta_i$ and $\hat\alpha_i$. Training minimizes the summed negative log likelihood of WFPT over trials. Interpretation reveals that subject-specific gamma band power (≈30–50 Hz) encodes evidence accumulation speed (drift), while beta band activity (≈20–30 Hz) around response encodes boundary/caution (Sun et al., 2022).

2. Neural Coding of Cognitive Load, Capacity, and Efficiency

Parametric fMRI analysis of working-memory load (n-back) under differential equation models provides voxel- and region-wise maps of neural efficiency (slope parameter $B/2$), neural capacity (peak activation $N_{cap} = -B^2/(4A) + C$), and cognitive capacity ($t^* = -B/(2A)$). The quadratic relationship

$x(t) = At^2 + Bt + C, \;\;\; dx/dt = 2At+B$

captures both linear and nonlinear dependencies of brain activity on task difficulty. Estimation via GLM with polynomial contrasts recovers interpretable maps: widespread frontoparietal and insular nodes are capacity-limited (A<0), with peak engagement at $t^* \approx 1.6$–$2.4$; neural efficiency correlates with behavioral accuracy and RT, supporting theories such as CRUNCH and sigmoidal resource models (Steffener et al., 2016).

Complementary EEG-based analyses focus on functional network organization: eigenvalue spectra of coherence matrices and graph-theoretic variability (degree, path length) predict moment-to-moment working-memory success above chance (AUC=0.63 for digit recall failure detection). Efficient neural organization—characterized by isotropic, “flat” eigenvalue structures and small-world topology—is linked to superior cognition, while collapse of such complexity marks failures (Helfer et al., 2016).

3. Network Generativity and Structural Constraints on Cognition

Generative models of the human connectome, such as action-based neural network generators, reconcile geometric and topological wiring principles to reproduce observed interindividual variability. In these models, the probability of forming a connection is

$P_{ij} \propto \alpha_{ij} \exp(-\eta d_{ij}),$

where $\alpha_{ij}$ encodes topological preference (e.g., degree-based, triadic closure) and $\eta$ controls the distance penalty. Subject-optimized $\eta_i$ provides a mechanistic index for the propensity to form long-range (“non-local”) connections. Empirically, lower $\eta_i$ (i.e., greater long-range wiring) predicts higher general intelligence ($r=-0.318$, $p=0.001$) and better cognitive test performance, linking global network integration to cognitive ability (Arora et al., 2022).

4. Latent Representations and Symbolic Computation in Neural Circuits

Population codes in cortical and hippocampal circuitry support temporally and spatially extended memory through Laplace-domain representations. Populations of leaky integrators encode the Laplace transform $F(s) = \int_0^\infty f(\tau) e^{-s\tau} d\tau$ of recent history, while inverse networks reconstruct a compressed but approximately scale-invariant timeline. These motifs generalize to abstract quantities (conceptual distances, numerosity) and enable population-level implementation of data-independent symbolic operators (translation, convolution, scaling), matching behavioral and single-unit evidence for compressed memory and flexible evidence accumulation (Howard et al., 2020).

5. Direct Cognitive–Neural Model Alignments: Decoding, Manipulations, and Dynamics

Modern approaches utilize representational similarity and encoding analyses to link neural and artificial representations:

• Vision–language DNNs (e.g. CLIP) align more closely with human VOTC activity than object-label- or unsupervised models, with fMRI-model fit strength depending on left-lateralized language pathway integrity. Lesions disrupting the VOTC–angular gyrus tract causally reduce CLIP’s explanatory power, indicating that language dynamically shapes visual cortical representations (Chen et al., 23 Jan 2025).
• Sequence models combining Hidden Multivariate Pattern (HMP) decomposition and structured state-space modeling (S4) decode trial-wise cognitive strategies, revealing the neural signature of an intermediate “Confirmation” operation in EEG. The probabilistic presence of this operation correlates with improved accuracy and change-of-mind behaviors, challenging homogeneous process assumptions in decision models (Otter et al., 14 Apr 2025).
• State-to-output controllability theory provides a framework for linking high-dimensional neural activity to low-dimensional manifolds organizing cognitive/affective states. Manifold structure and empirical controllability Gramian eigenspectra predict how readily cognitive interventions steer subjective experience, validated in both healthy fMRI and clinical depression symptom trajectories (Shariatpanahi et al., 18 Mar 2025).

6. Neural Evidence in Artificial and Hybrid Cognitive Systems

Neural networks’ cognition and knowledge representation can be analyzed at the neuron level through epistemological markers: each neuron encodes conceptual axes (“concepts-in-action”), and their weight vectors constitute “theorems-in-action,” equivalent to micro-theories. Cognitive operations (selection, fusion, abstraction) are realized via weighted aggregation and interlayer transformation, formalized as

$X^{(\ell)} = W^{(\ell)} X^{(\ell-1)},$

building hierarchies of meaning across layers. Systematic probing via ablation, basis projection, and activation-classification grounds the claim that artificial networks realize computationally coherent cognitive states (Pichat, 4 Jul 2024).

Empirical evaluation frameworks contrast neurobiological encoding performance of neural LLMs (transformers, word embeddings) and psychologically-plausible models (embodied ratings, topological graphs). In bilingual fMRI and eye-tracking, embodied and topological models, leveraging experiential feature ratings and word-graph properties, outperform even large neural LLMs in predictive accuracy, particularly in core semantic regions such as IFG, AG, and MTG. This establishes strong evidence for the sufficiency (and sometimes superiority) of structured, psychologically-plausible representations relative to large data-driven networks (Zhang et al., 30 Apr 2024).

7. Implications, Limitations, and Future Directions

The convergence of behavioral, neural, and computational evidence substantiates multi-level models wherein cognitive processes are instantiated in dynamically organized brain networks. Quantitative benchmarks—likelihood fits, spectral saliency, encoding/decoding performance, statistical associations with structural parameters—anchor theoretical constructs in observable neurodynamics.

Cognitive-neural evidence is inherently multi-modal, requiring joint consideration of behavior, brain activity (EEG, fMRI), network structure, and computational form. While single-trial and subject-specific modeling reveal granularity previously masked by averaging, current limitations include reliance on limited tasks, small cohorts, and indirect proxies for certain latent variables. Extensions to more ecologically valid tasks, richer neural recording regimes, and causal manipulations (lesions, stimulation) will refine these models further, deepening our mechanistic grasp of cognition as implemented in neural systems.