Third Visual Pathway in Primate Vision

• Third visual pathway is a distinct anatomical channel linking LGN, V5/MT, and STS, specialized for processing dynamic social cues and rapid motion.
• Advanced neuroimaging, lesion analysis, and computational modeling reveal its unique circuit architecture and functional tuning distinct from classical streams.
• Clinical studies in dyslexia and comparative DNN analyses underscore its relevance and challenge existing paradigms in visual computation.

The third visual pathway is a major anatomical and functional substrate in the primate (including human) visual system, distinguished from the classical ventral (“what”) and dorsal (“where/how”) streams. This pathway comprises direct projections that bypass canonical relays, linking early visual, thalamic, and higher-order cortical loci, and subserves dynamic social perception, rapid motion processing, and the binding of visual information with spatial and social context. Recent research integrating anatomical tracing, ultra-high-field MRI, neuroimaging, neuropsychological lesion analysis, and computational modeling has converged on the third visual pathway as a functionally discrete, evolutionarily specialized channel, with broad implications for theories of visual computation, social cognition, and neural network architecture (Müller-Axt et al., 2017, Marvi et al., 9 Oct 2025, Pitcher, 10 Dec 2025, O'Reilly et al., 2017).

1. Anatomical Organization and Circuitry

The third visual pathway—also referred to as the lateral pathway—demonstrates a distinct trajectory from both the ventral (V1→V2→V4→IT) and dorsal (V1→V2→V3→PPC) streams. Multiple lines of anatomical evidence delineate this pathway:

• Thalamo-cortical Branch: Non-human and human studies have demonstrated a direct projection from the lateral geniculate nucleus (LGN), especially its koniocellular layers, to area V5/MT, bypassing primary visual cortex (V1). This tract runs dorsal to the geniculostriate bundle and can be noninvasively reconstructed in vivo using diffusion MRI and probabilistic tractography (Müller-Axt et al., 2017).
• Cortico-cortical Progression: In primates, the pathway extends polysynaptically as V1 → V5/MT → MST → FST → posterior STS (pSTS), with each link exhibiting strong feedforward and local recurrent connectivity (Pitcher, 10 Dec 2025).
• White-matter Fasciculi: Human diffusion-tractography localizes the “lateral fasciculus” (running in the fundus of STS), clearly separable from the inferior longitudinal (ILF) and inferior fronto-occipital fasciculus (IFOF) of the ventral stream and the superior longitudinal fasciculus (SLF) of the dorsal stream. Quantified strengths for the right middle longitudinal fasciculus (MdLF: $N_{\rm MdLF}=3200\pm550$, FA$_{\rm MdLF}=0.42\pm0.05$) and arcuate fasciculus (AF: $N_{\rm AF}=2800\pm450$, FA$_{\rm AF}=0.44\pm0.04$) confirm their major role in lateral stream integrity (Pitcher, 10 Dec 2025).
• Functional Neuroanatomy: The pathway terminates in lateral occipito-temporal cortex—specifically pMTG, pSTS, and adjacent lateral occipital sulcus—distinct from classical ventral and dorsal categorical or motion/action patches (Marvi et al., 9 Oct 2025).

Stream	Major Nodes	Distinguishing White-Matter Tracts
Ventral	V1 → V2 → V4 → IT/FFA	ILF, IFOF
Dorsal	V1 → V2 → V3 → MT → PPC	SLF, parietal projections
Third (Lateral)	LGN → V5/MT → MST → STS	MdLF, AF (“lateral fasciculus”)

2. Functional and Computational Specialization

The third visual pathway displays unique computational tuning and regional selectivity:

• Dynamic Social Cues: fMRI localizes pSTS as exhibiting strong BOLD selectivity for dynamic over static faces ($\Delta$BOLD$\approx$+0.30%, $p<0.001$), as well as for moving bodies and audiovisual speech integration (Pitcher, 10 Dec 2025). Lesion studies confirm that pSTS damage selectively impairs dynamic but not static expression recognition, whereas ventral lesions show the converse dissociation.
• Component Tuning Profiles: Sparse decomposition of fMRI responses reveals lateral components for group interactions, implied motion, hand actions, scenes, and reachspaces, with specific behavioral axes (motion, social, hand-action saliency) correlating with neural tuning: corr($R_{·,\text{implied motion}}$, motion ratings)$\,\approx0.66$ (Marvi et al., 9 Oct 2025).
• Receptive Field and Temporal Properties: Lateral stream regions exhibit large, bilateral receptive fields (≤60°), no hemifield bias, and fast motion selectivity (MT onset ~40 ms; pSTS 60–140 ms). Compared to the ventral face/form network, which operates over small, contralateral fields with no dynamic preference and has distinct latency profiles, the third pathway is specialized for panoramic, rapid decoding of time-varying, socially relevant information (Pitcher, 10 Dec 2025).
• Computational Models: Models of pSTS neurons as spatiotemporal motion-energy filters capture their velocity/orientation sensitivity; recurrent intention-decoding architectures support the integration of motion cues with social inference, with dynamics governed by

$$\tau\frac{d\mathbf{r}(t)}{dt} =-\,\mathbf{r}(t)+W\,\mathbf{r}(t)+U\,\mathbf{E}(t)$$

and subsequent softmax readouts over intention templates (Pitcher, 10 Dec 2025).

3. Structural and Clinical Relevance

In vivo imaging in neurotypical and clinical populations has elucidated the behavioral importance of the third visual pathway:

• Dyslexia and Reading: Ultra-high-field MRI and probabilistic tractography in adults with developmental dyslexia reveal a selective reduction of the direct LGN–V5/MT pathway in the left hemisphere, with connectivity index $I$ significantly lower in dyslexics ($I\approx0.68\pm0.02$) than controls ($I\approx0.72\pm0.02$), $t(22)=3.13$, $P=.005$ (Müller-Axt et al., 2017). Critically, left V5/MT–LGN connectivity strength correlates with rapid automatized naming speed ($r=-0.588$, $P=0.045$), a key behavioral marker of dyslexia, while no differences are found for the canonical LGN–V1 projection.
• Neuropsychological Lesions: STS lesions in humans cause selective impairment for dynamic social cues, measured as $d'_{\text{dynamic,STS}}=0.45\pm0.08$, but not for static cues ($d'_{\text{static,STS}}=1.20\pm0.10$), while the reverse holds for ventral lesions (Pitcher, 10 Dec 2025). Prosopagnosic patients with ventral face-selective area damage retain pSTS activation to dynamic faces.
• Subcortical–Cortical Integration: These findings emphasize the role of subcortical–cortical loops, expanding traditional cortico-centric models of visual dysfunction (e.g., in dyslexia) to include thalamo-cortical pathways with modality-specific deficits (Müller-Axt et al., 2017).

4. Hierarchical Visual Models and Cross-Stream Integration

The computational architecture and development of the third visual pathway can be contextualized within models positing three distinct but interactive visual streams:

• Integration in the What × Where Model: The “What × Where” (Editor’s term: WWI) stream (V2 → V3/MT → V4/TEO) learns conjunctive spatio-feature representations, binding object features with spatial and motion information to generate predictive signals. The unique laminar and hierarchical structure, involving deep-layer corticothalamic projections and temporally partitioned error signals broadcast via the pulvinar, supports error-driven learning and abstraction (O'Reilly et al., 2017).
• Error Partitioning Mechanism: In this framework, the third stream absorbs the residual “mixed” error (feature-location conjunctions) after pure spatial and object-identity errors are factored out by dorsal and ventral pathways. Layer-specific learning is governed by local Hebbian update rules such as

$$\Delta w_{ij} \;\approx\; \eta\,\bigl(x_i^+\,y_j^+ - x_i^-\,y_j^-\bigr)$$

and more generally by the XCAL function for error-driven and homeostatic adjustment (O'Reilly et al., 2017).

5. Comparative Modeling, Alignment, and Implications for Artificial Systems

High-resolution decomposition and cross-system alignment reveal distinct computational signatures of the third visual pathway:

• Sparse Component Alignment (SCA): This methodological advance measures alignment between neural representational axes (as extracted via Bayesian non-negative matrix factorization) and artificial neural networks. SCA is sensitive to native system axes, in contrast to conventional RSA. It shows that feedforward DNNs trained on single-image tasks align strongly with ventral stream organization (SCA $r=0.187$ for AlexNet), but nearly fail to capture the lateral stream (SCA $r=0.047$) (Marvi et al., 9 Oct 2025).
• Functional Dissociation in Models: Current DNNs fail to recover unique lateral stream tuning (specialized for social and implied-motion cue processing), indicating that standard image recognition objectives and architectures are insufficient for capturing this specialized stream.
• Future Modeling Directions: Accurate modeling of the third pathway may require architectures trained on dynamic, temporally extended and social interaction tasks, explicit sparsity constraints, and specialized local readouts replicating sparse component axes (Marvi et al., 9 Oct 2025). SCA offers a principled tool for neurocomputational benchmarking.

Metric	Ventral	Lateral	Dorsal
Linear Encoding (r)	0.180	0.179	0.232
Representational Similarity (RSA, ρ)	0.347	0.222	0.199
Sparse Component Alignment (r)	0.187	0.047	0.058

6. Broader Significance and Future Directions

The elucidation of the third visual pathway establishes a new principal axis in visual cortical organization, with distinctive roles in:

• Social Perception: Rapid decoding of facial expressions, gaze, biological motion, action intent, and other social signals, critical for human social cognition and communication (Pitcher, 10 Dec 2025).
• Neurodevelopment: Developmental imaging suggests this pathway emerges early, possibly supporting the early acquisition of social visual skills.
• Pathology and Intervention: Selective deficits (e.g., in dyslexia or after STS lesions) motivate intervention strategies aimed at strengthening specific subcortical–cortical circuits.
• Comparative Neuroscience: Cross-species tract-tracing and homology mapping will refine understanding of the phylogenetic emergence and specialization of the pathway.
• Methodological Advances: High-resolution laminar fMRI, optogenetic perturbation, and sparsity-based analysis in both model systems and human cortex will further dissect the causal and representational basis of the third pathway.

In summary, the third visual pathway—spanning thalamo-cortical and cortico-cortical projections from LGN and early visual cortex through V5/MT to STS and lateral occipito-temporal cortex—serves specialized processing of dynamic, social, and conjunction-rich visual inputs. Its anatomical distinctness, computational specialization, clinical significance, and failure of current artificial models to replicate its representational axes define it as a central topic in contemporary visual neuroscience (Müller-Axt et al., 2017, Marvi et al., 9 Oct 2025, Pitcher, 10 Dec 2025, O'Reilly et al., 2017).