Locating Plasticity in the Sensory Decision Hierarchy

30 concepts20 papers1 proposal

When we practice a sensory task—such as distinguishing subtle shades of contrast or isolating a specific sound from a noisy background—our performance improves dramatically over time. This phenomenon, known as perceptual learning, raises a fundamental question about the functional architecture of the brain: where does this adaptation actually happen?

Does the brain rewire its early sensory "cameras" (like the primary visual or auditory cortices) to capture finer details? Or do early representations remain fixed while downstream "analysts" (higher-order decision and readout areas) learn to interpret the incoming data more efficiently?

For decades, neuroscientists have faced conflicting empirical evidence. Some experimental paradigms point to early sensory refinement, while others suggest that learning is strictly a downstream phenomenon. Recent work reconciles these views by framing perceptual learning not as a single localized change, but as a coordinated geometric optimization across the sensory-decision hierarchy.

The Case for Early Sensory Plasticity

One school of thought argues that practice fundamentally reshapes early sensory processing. When a task demands fine-grained discrimination, early cortical areas appear to warp their representations to prioritize behaviorally relevant features.

For instance, Ley et al. (2012) demonstrated that learning new sound categories causes the primary auditory cortex to shift from encoding raw physical features to mapping abstract categorical boundaries. A comprehensive review of auditory abstraction [How learning to abstract shapes neural sound representations] similarly highlights how experience physically warps early perceptual space to emphasize behaviorally relevant information over raw sensory data.

This early-stage plasticity is heavily driven by top-down modulation. In studies of associative learning and emotional conditioning, Chen et al. (2019) and Ji et al. (2023) showed that top-down signals actively reorganize low-level sensory cortices, creating stable, discriminative response patterns where stimulus-driven processing normally dominates.

In the visual domain, Shibata et al. (2012) pinpointed cortical area V3A as a primary site of sensory refinement for motion learning, showing a tight correlation between behavioral improvement and changes in neural tuning functions. Even social context can trigger this low-level reorganization: Paraouty (2023) found that observing a peer perform a task accelerates an animal's own subsequent learning by inducing direct plasticity in the observer's auditory cortex.

The Case for Downstream Readout Refinement

Conversely, a compelling body of work suggests that early sensory filters remain stubbornly fixed, and learning is instead a matter of optimizing how downstream decision centers "read out" this stable information.

In a classic study on orientation discrimination, Ghose et al. (2002) found that despite extensive training, the basic response properties of neurons in early visual areas (V1 and V2) were virtually unchanged. The authors argued that behavioral gains must instead stem from a more sophisticated pooling or interpretation of these stable signals downstream. Supporting this, Lowe et al. (2013) tracked contrast learning and found that while information in V1 remained static, behavioral improvements correlated precisely with spiking activity in the mid-tier visual area V4.

This pattern repeats across different sensory modalities and complexities:

  • Multisensory Calibration: Zaidel et al. (2021) observed neural shifts reflecting behavioral adaptation in the higher-order ventral intraparietal area (VIP), while the earlier dorsal medial superior temporal area (MSTd) remained unchanged.
  • Fine Contrast Discrimination: Thiele et al. (2024) demonstrated that contrast training yields only modest, highly task-dependent changes in V1, reinforcing the idea that higher-level decision centers drive the perceptual "sharpening."
  • Subcortical Dissociations: When examining patients with damaged basal ganglia, Lu et al. (2010) found that while category learning was impaired, perceptual learning in low-noise environments was spared, pointing to distinct, distributed neural pathways for handling different environmental noise conditions.

Reconciling the Hierarchy via Neural Geometry

How can both perspectives be true? Computational and empirical work suggests that perceptual learning is not a single-locus event but a coordinated, multi-stage process across the entire cortical hierarchy. As Xi et al. (2020) showed using event-related potentials, contrast learning triggers a wave of plasticity that spans from the earliest sensory stages to high-level decision areas, making processing both faster and more sensitive at every step.

The bridge between early representation and downstream readout lies in how information is structured. Panniello et al. (2024) found that during learning, the primary somatosensory cortex increases its sensory information, allowing downstream regions to process and use that information more effectively to guide choices. However, modeling this process has historically been difficult; early computational frameworks struggled to definitively distinguish whether the brain prioritizes signals via noise filtering or signal accommodation, as noted by Sundareswaran et al. (1994).

A recent breakthrough by Cheng et al. (2025) reconciles these conflicting models through the lens of the geometry of neural populations. Rather than focusing solely on individual neuron tuning or population noise correlations—a long-standing debate highlighted by Sanayei et al. (2018)—the authors demonstrate that the primary driver of visual perceptual learning is manifold shrinkage.

Before Learning: High Variability          After Learning: Manifold Shrinkage
       (Overlapping Responses)                    (Clean Separation)

          Task A     Task B                          Task A     Task B
         (  *   )   (   *  )                        ( * )      ( * )
        ( *  *   ) (   *  * )                        (*)        (*)
         (  *   )   (   *  )                          *          *
            \           /                               \      /
             \         /                                 \    /
          Downstream Readout                          Downstream Readout
          Struggles to Decode                         Easily Decodes Boundary

By treating the collective response of a neural population as a geometric manifold in a high-dimensional space, Cheng et al. (2025) showed that learning globally reduces trial-by-trial response variability. This shrinkage makes the population representation cleaner and more compact. Consequently, downstream decision centers can read out the information with far greater accuracy without requiring a radical rewiring of individual sensory filters.

This optimization comes with a fascinating trade-off. As modeled by Szpiro et al. (2022), sharpening population representations to maximize discrimination accuracy near a decision boundary can paradoxically distort the subjective appearance of a stimulus. The brain becomes better at telling subtle differences apart near the trained boundary, but it does so at the cost of a more biased, distorted perception of stimulus magnitude overall.

Ultimately, perceptual learning works because the brain optimizes the interaction between stages: early sensory cortices compress their response manifolds to reduce noise, making it trivial for downstream decision centers to implement a more efficient readout strategy.

Key Foundational Work

Position Variance, Recurrence and Perceptual Learning (2000) — contributed 1 concept elements to this direction.
Learning direction in global motion: two classes of psychophysically-motivated models (1994) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_cortical_plasticity_localization_84e159d8,
  title = {cortical plasticity localization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cortical-plasticity-localization/rdr_84e159d88de14ce69e57b56914ffc757},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_cortical_plasticity_localization_84e159d8_v1,
  title = {cortical plasticity localization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cortical-plasticity-localization/rdr_84e159d88de14ce69e57b56914ffc757/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake