NIH-funded study proposes blurry vision in babies may guide brain development

Newswise — Shortly after birth when the world is a blur, babies may be learning to identify patterns. According to a new study funded by the National Eye Institute (NEI), the initial phase of blurry vision may be fundamental to the development of normal visual processing. The study appeared online October 15 in the Proceedings of the National Academy of Sciences. The NEI is part of the National Institutes of Health.

“The findings inform our understanding of how normal vision develops, which is important for addressing disorders that affect the brain’s visual processing system,” said Cheri Wiggs, Ph.D., program director for low vision and blindness rehabilitation at the NEI.

At birth, the shape of the eye does not allow images to be seen in sharp focus. As babies grow, their eyes change shape, improving visual focus. At the same time, areas of their brains that process visual images develop, forming connections and networks that help the infants recognize objects.

Children who never experience this initial phase of out of focus imagery can later have difficulty with complex visual tasks.

Such is the case for people born blind from cataracts who undergo corrective surgery later in childhood. Once treated, their visual acuity (ability to read letters on an eye chart) immediately improves.

But they’ve been observed to have long-term difficulty with facial recognition, said the study’s senior investigator Pawan Sinha, Ph.D., professor of vision and computational neuroscience at Massachusetts Institute of Technology.

“The rapid attainment of relatively high visual acuity after cataract surgery may be an important factor that pushes a child’s developmental trajectory away from the normal one, wherein acuity improves gradually after initially being quite poor,” he said.

Sinha and colleagues tested what they dubbed the “initial impoverishment benefit hypothesis” using a deep convolutional neural network—a computer-based image classification system with some similarities to the biological one. Much like the way a large number of interconnected neurons are involved in visual processing, deep neural networks rely on a constellation of many simple computing units whose connections to each other can strengthen or weaken with experience. This modifiability of connections helps them learn from errors and seek adaptations that improve classification accuracy over time. The application of such artificial intelligence systems is being explored in many fields including the detection of abnormalities on diagnostic and screening imaging tests.

The researchers trained deep neural networks on a large database of facial images. They simulated three different visual training scenarios by systematically blurring and increasing the resolution of the images. In the first, they replicated an infant’s normal course of vision by training the network on blurry images, followed by high-resolution ones. In the second, the training order was reversed; high-resolution images were followed by low-resolution images. And lastly, the network was trained on only high-resolution images.

At any point in the training, the introduction of blurred images automatically induced the neural network to expand the size of its receptive fields, the spatial areas detectable by the network’s individual visual sensors. In other words, blurry images lacked detail, thereby prompting the neural network to seek more information by enlarging the size of the receptive fields. This information integration across larger areas improved the network’s ability to recognize patterns, potentially by enabling it to use more global image cues. Once this adaptation was acquired, the neural network retained it. The introduction of high-resolution images following the blurred ones, for example, did not shrink the size of the receptive fields. And even when the period of blurred image training was relatively short, it was sufficient to induce enlargement of the receptive fields.

Larger receptive fields proved key for image recognition. The neural network performed optimally when training initially involved blurry images followed by high-resolution ones. By contrast performance was poor when training initially involved high-resolution images followed by blurry ones, even though the network had been trained with precisely the same set of images as with the blurry to high-resolution training. Performance was poorer still when training had been based on high-resolution images only.

“Caution needs to be exercised when using such neural networks as proxies of the human visual system,” Sinha said. There may be important differences between the two systems. “However, the initial stages of the deep networks, which are what we focused on in this study, do show striking similarities with their biological counterparts in terms of their response properties. This makes us more confident about using these stages of deep networks as approximate models of early visual processes in the mammalian brain.”

“From a practical perspective, these findings suggest a superior strategy for training deep neural networks,” he added. “Training neural networks on high resolution images has been a common default. Perhaps image recognition could be improved by drawing inspiration from human development, and training networks first with degraded images.”

The findings also suggest that children who have surgery for cataracts may benefit from a more gradual improvement in their visual acuity.

Future studies are needed to explore the initial impoverishment benefit hypothesis in other domains. For example, an analogous notion may also apply to auditory stimuli. Fetuses in the womb hear environmental sounds muffled by the amniotic medium. It could be that newborns born prematurely would benefit from an extended period of similarly muffled sounds to mimic the early auditory experiences of a full-term infant.

The study was funded by NEI grant EYR01020517.

Reference:

Vogelsang L, Gilad-Gutnick S, Ehrenberg E, Yonas A, Diamond S, Held R, Sinha P. Potential downside of high initial visual acuity. PNAS October 30, 2018 115 (44) 11333-11338; published ahead of print October 15, 2018 https://doi.org/10.1073/pnas.1800901115

###

This press release describes a basic research finding. Basic research increases our understanding of human behavior and biology, which is foundational to advancing new and better ways to prevent, diagnose, and treat disease. Science is an unpredictable and incremental process— each research advance builds on past discoveries, often in unexpected ways. Most clinical advances would not be possible without the knowledge of fundamental basic research.

NEI leads the federal government’s research on the visual system and eye diseases. NEI supports basic and clinical science programs to develop sight-saving treatments and address special needs of people with vision loss. For more information, visit https://www.nei.nih.gov.

About the National Institutes of Health (NIH): NIH, the nation’s medical research agency, includes 27 Institutes and Centers and is a component of the U.S. Department of Health and Human Services. NIH is the primary federal agency conducting and supporting basic, clinical, and translational medical research, and is investigating the causes, treatments, and cures for both common and rare diseases. For more information about NIH and its programs, visit https://www.nih.gov/.

NIH…Turning Discovery Into Health®