Summary: New research from MIT shows how the poor, low-color vision experienced in early infancy can shape the brain’s visual pathways into specialized systems. Using computational models trained first on low-quality images and later on high-quality ones, the team demonstrated that this developmental progression encourages the emergence of processing units resembling magnocellular and parvocellular channels.
The findings imply that initial sensory limitations—blurred input and reduced color sensitivity—may help the visual system build robust abilities to perceive shape, motion, and fine detail later in life. The study sheds light on how visual experience during development interacts with the organization of the brain.
Key Facts:
- Early exposure to blurry, grayscale visual input helps development of specialized visual pathways.
- Computational models trained on biomimetic input developed magnocellular- and parvocellular-like processing units.
- Biomimetic training produced models that relied more on shape than texture when recognizing objects, resembling human strategies.
Source: MIT
Incoming information from the retina is relayed into two main pathways in the brain’s visual system: one tuned to color and fine spatial detail and another tuned to spatial location and rapid changes over time.
A new MIT study offers a developmental explanation for how these two pathways might arise.
Newborns have limited visual acuity and weak color perception because their retinal cone cells are not fully developed at birth. As a result, infants’ early visual experience is dominated by blurred, desaturated images.
The MIT researchers propose that these early conditions encourage a subset of visual neurons to specialize in low spatial frequencies and reduced color sensitivity—attributes associated with the magnocellular system. As the visual system matures and image quality improves, other neurons become tuned to fine detail and richer color information, consistent with the parvocellular pathway.
To test this idea, the team trained deep computational models on a developmental sequence that mimics infant visual input: an initial phase of low-resolution, grayscale images followed by a later phase of high-resolution, full-color images.
Models trained on this biomimetic schedule developed processing units with receptive-field properties that resemble the parvo/magno division. By contrast, models trained only on high-quality, full-color images from the start did not develop such distinct specializations.
“These findings offer a potential mechanistic account for the parvo/magno distinction, a major organizing principle of the mammalian visual pathway,” says Pawan Sinha, professor of brain and cognitive sciences at MIT and senior author of the study.
The paper’s lead authors are MIT postdocs Marin Vogelsang and Lukas Vogelsang. Coauthors include MIT research affiliate Sidney Diamond and Gordon Pipa, a professor of neuroinformatics at the University of Osnabrueck. The study is published in Communications Biology.
Sensory input
This line of research was inspired in part by Project Prakash, an effort led by Sinha’s lab that screens and treats children with reversible forms of blindness in India. Studies of children who regain sight after congenital cataracts reveal how visual abilities develop when sensory input is restored.
One Project Prakash finding showed that children whose cataracts were removed performed worse on object-recognition tests when shown black-and-white images compared with color images. That result suggested that early reduced color input might paradoxically help the visual system learn to recognize objects even when color information is unreliable.
In computational experiments, models first trained on grayscale images and later exposed to color images became more resilient to color changes than models trained only on color images. Similar effects were observed when models experienced a progression from blurry to sharp images.
Building on those results, the MIT team investigated whether simultaneous developmental limitations in spatial resolution and chromatic sensitivity could underlie the parvocellular–magnocellular distinction.
Parvocellular neurons are typically sensitive to color and have small receptive fields, letting them resolve fine detail. Magnocellular neurons, in contrast, pool signals across broader retinal areas and are sensitive to motion and rapid changes, emphasizing global spatial information and temporal dynamics.
The researchers compared models trained on a standard image dataset with models trained on a biomimetic dataset that simulated early human input: a first training phase of low-resolution, grayscale images, followed by a second phase of high-resolution, colorful images.
After training, the team examined individual processing units within the networks—analogous to groups of neurons—to determine their spatial and chromatic tuning. Models trained on biomimetic input developed a coherent subset of units responsive to low color and low spatial frequencies, consistent with magnocellular-like tuning. Other units displayed higher spatial-frequency or stronger color tuning, resembling parvocellular-like properties. This clear division did not arise when training began with high-quality, full-color images.
“These results support the idea that correlations seen in the biological visual system could reflect the particular combinations of inputs available during normal development,” Lukas Vogelsang explains.
Object recognition
To understand how these different training histories affect recognition strategies, the researchers tested the models on images where shape and texture were mismatched—such as an object with the silhouette of one animal but the surface texture of another. This approach reveals whether models rely on global shape or local texture to categorize objects.
Models trained with biomimetic input relied more heavily on global shape, paralleling human behavior. Removing magnocellular-like units from those networks reduced their shape-based recognition, indicating that magnocellular-like processing supports shape-driven categorization.
The team also trained models on video data to introduce temporal dynamics. Magnocellular pathways are known to respond to high temporal frequencies, helping detect rapid changes in position. In biomimetically trained video models, units tuned to high temporal frequencies also tended to show magnocellular-like spatial properties.
Together, these results indicate that low-quality sensory input early in life can contribute to the emergence and organization of specialized visual pathways. While the findings do not exclude innate biological factors, they demonstrate that developmental visual experience can play a meaningful role.
“It appears that the staged progression of visual input during development is carefully structured to produce particular perceptual strengths and may even shape the brain’s organizational layout,” Sinha says.
Funding:
This research was supported by the National Institutes of Health, the Simons Center for the Social Brain, the Japan Society for the Promotion of Science, and the Yamada Science Foundation.
About this visual neuroscience and neurodevelopment research news
Author: Sarah McDonnell
Source: MIT
Contact: Sarah McDonnell – MIT
Image: The image is credited to Neuroscience News
Original Research: Open access. “Potential role of developmental experience in the emergence of the parvo-magno distinction” by Pawan Sinha et al., Communications Biology
Abstract
Potential role of developmental experience in the emergence of the parvo-magno distinction
The division of the early visual pathway into parvocellular and magnocellular systems, each with distinct response properties, is a well-established organizing principle in mammalian vision. However, the developmental origins of this division are not fully understood.
This study proposes that the overlapping developmental timelines of spatial frequency sensitivity and chromatic sensitivity can shape the emergence of neuronal response properties characteristic of parvo- and magnocellular pathways. Analyses of receptive fields in deep networks trained with developmentally inspired “biomimetic” protocols support this proposal in both spatial and temporal domains.
Biomimetic training also produced a stronger, more human-like bias toward global shape processing, potentially driven by magnocellular-like units. These results have implications for understanding the development of a key aspect of visual pathway organization and for designing training regimes for artificial vision systems.