For decades, textbook neuroscience described sensory processing as a strict, serial hierarchy. In vision, for instance, photons hit the retina, signals travel to the lateral geniculate nucleus (LGN) of the thalamus, pass to the primary visual cortex (V1), and then march sequentially through upstream areas to build complex representations from simple edges.
However, this assembly-line model introduces a severe vulnerability: a single break in the chain should collapse the entire system. Yet, primates with damaged primary visual cortices can still navigate obstacles—a phenomenon known as blindsight. As reviewed by Alessia Celeghin et al. in Convolutional neural networks for vision neuroscience: significance, developments, and outstanding issues, traditional feedforward computational models fail to account for these residual capabilities because they ignore non-hierarchical subcortical pathways.
The brain does not rely on a single serial pipeline. Instead, it operates via a highly parallel, multi-branch architecture where subcortical structures and direct thalamocortical shortcuts process information concurrently.
Subcortical Autonomy and Thalamic Shortcuts
Why route sensory data through multiple pathways at once? The primary reason is functional specialization without temporal delay. Rather than serving as passive relay stations that merely pass raw data up to the "smart" cortex, subcortical structures act as independent processing hubs capable of integrating complex behavioral contexts.
Consider how the brain handles sound. Traditional models assume that integrating acoustic stimuli with behavioral goals happens exclusively in the cortex. However, Tai-Ying Lee et al. demonstrated in Midbrain encodes sound detection behavior without auditory cortex that the midbrain can encode sound detection behavior even when the auditory cortex is entirely absent.
In the visual domain, Peter Zeidman highlights in Distinct Top-down and Bottom-up Brain Connectivity During Visual Perception and Imagery that models assuming input enters exclusively through primary visual areas overlook parallel subcortical pathways like the superior colliculus, which directly balances bottom-up and top-down signals. This autonomy explains why higher-order cortical areas can possess sensory signals that are not inherited from their primary hierarchical inputs. For instance, Syed A. Chowdhury et al. discovered that the Middle Temporal (MT) area carries vestibular signals related to self-motion even when its primary upstream visual inputs lack those signals, implying independent subcortical or cerebellar pathways (Does the Middle Temporal Area Carry Vestibular Signals Related to Self-Motion?).
Traditional Serial Model:
Retina ──> Thalamus ──> Primary Cortex (V1) ──> Secondary Cortex ──> Higher Behavior
Parallel Non-Serial Model:
┌──> Thalamus ───────────────> Secondary Cortex (Direct Shortcut)
Retina ───┼──> Subcortical Hubs (SC) ──> Higher Behavior (Rapid Response)
└──> Thalamus ──> V1 ────────> Deeper Cortical Layers
Furthermore, the thalamus itself bypasses primary sensory cortices to feed secondary areas directly. In the somatosensory system, Sami El‐Boustani et al. discovered that secondary whisker cortices receive direct, high-fidelity sensory information from the thalamus rather than relying on hierarchical input from primary areas (Anatomically and functionally distinct thalamocortical inputs to primary and secondary mouse whisker somatosensory cortices).
This direct routing also occurs within the cortical layers themselves: Kevin Kermani Nejad et al. showed that deeper cortical layers receive direct, parallel thalamic inputs, challenging the sequential layer-to-layer flow model (Self-supervised predictive learning accounts for cortical layer-specificity). Even complex cells in the visual cortex, as modeled by Alexander Grunewald et al., likely receive direct thalamic inputs to detect second-order motion, bypassing simple cells entirely (Detection of First and Second Order Motion).
Temporal Concurrency and Feature Integration
If the brain were strictly serial, we would perceive an object's low-level features (like color or orientation) before understanding its semantic category. But empirical evidence shows that the brain extracts these features simultaneously.
Monica Betta et al. tracked object recognition and found that low-level features, geometric shapes, and semantic categories are processed concurrently within a shared temporal window (Common spatiotemporal processing of visual features shapes object representation). This temporal concurrency is mirrored in higher-order visual areas during active tasks. For instance, Noam Roth et al. observed that goal-related target signals can appear in high-level regions like the inferior temporal cortex simultaneously with or even before lower-level regions like V4, indicating a non-hierarchical distribution of top-down instructions (The integration of visual and target signals in V4 and IT during visual object search).
This parallel extraction applies across modalities and feature dimensions:
- Magnitude and Duration: Alessandro Toso et al. demonstrated that tactile intensity and duration are computed concurrently from the same neural input stream via a parallel integration architecture, allowing post-hoc retrieval of either feature without performance degradation (A sensory integration account for time perception).
- Existence and Characteristics: In cognitive bottlenecks like the attentional blink, Swagata Halder et al. showed that detecting a stimulus's existence and discriminating its specific features are handled by concurrent, parallel computations rather than a sequential pipeline (Distinct neural bases of subcomponents of the attentional blink).
- Cortical Fields: In the auditory cortex, Jennifer K. Bizley et al. found that feature sensitivity is distributed across multiple cortical fields in parallel, rather than being confined to a single specialized region (Neural Ensemble Codes for Stimulus Periodicity in Auditory Cortex).
Why Parallelism Works: Computational Advantages
Why did evolution favor this multi-branch, non-serial architecture? Parallel sensory pathways solve three fundamental computational challenges: speed, resilience, and the simultaneous management of conflicting demands.
1. Speed and Direct Action Pathways
Parallel pathways enable rapid behavioral responses. By bypassing the slow cortical hierarchy, sensory areas can directly drive motor centers. As noted by Emmanouil Froudarakis et al., the primary visual cortex can project directly to subcortical motor centers to execute rapid orienting responses and eye movements based on immediate sensory input (The Visual Cortex in Context).
2. Resilience and Flexible Negotiation
Parallel architectures provide resilience against sensory mismatches. For example, our internal spatial mapping relies on semi-independent information streams. Guifen Chen et al. showed that place cells (prioritizing visual landmarks) and grid cells (tethered to physical movement) function as parallel streams, allowing the brain to flexibly negotiate between different sensory modalities rather than relying on a single downstream coordinate system (Differential influences of environment and self-motion on place and grid cell firing).
When early-stage parallel circuits are disrupted—such as the breakdown of binocular integration in the thalamus during monocular deprivation studied by Carey Y. L. Huh et al.—the downstream deficits become permanent precisely because the hierarchy inherits a corrupted parallel stream (Long-term Monocular Deprivation during Juvenile Critical Period Disrupts Binocular Integration in Mouse Visual Thalamus).
3. Concurrent Integration and Segregation
Parallel processing allows the brain to perform seemingly contradictory operations at the same time, such as integrating and segregating multisensory cues. Wenhao Zhang et al. ("Congruent" and "Opposite" Neurons: Sisters for Multisensory Integration and Segregation) and He Wang et al. (Complementary congruent and opposite neurons achieve concurrent multisensory integration and segregation) demonstrated that "congruent" and "opposite" neurons work in tandem across distributed networks. This allows the brain to simultaneously combine related cues (integration) while detecting cue divergence (segregation).
Similarly, Adrien Combaz showed that the brain can process distinct visual stimuli for event-related potentials and frequency-based responses in parallel without signal degradation or resource competition (Simultaneous Detection of P300 and Steady-State Visually Evoked Potentials for Hybrid Brain-Computer Interface).
Ultimately, this multi-branch reality is reshaping how we build artificial intelligence. Liwei Huang et al. found that deep spiking neural networks with multi-branch structures approximate the biological activity of the visual cortex far better than purely serial architectures (Deep Spiking Neural Networks with High Representation Similarity Model Visual Pathways of Macaque and Mouse). By moving away from strict serial assembly lines and embracing parallel, non-hierarchical pathways, both neuroscience and AI are uncovering a more flexible, resilient, and rapid framework for sensory integration.
Key Foundational Work
Detection of First and Second Order Motion (1997) — contributed 1 concept elements to this direction.
Does the Middle Temporal Area Carry Vestibular Signals Related to Self-Motion? (2009) — contributed 1 concept elements to this direction.