One powerful infant behaviour that can be leveraged in developmental research is infants’ looking patterns. Studies of infants’ ability to group visual stimuli have revealed that young infants are sensitive to things like outlines, contrasts or shapes that can be used to form distinct categories such as cat and dog (Quinn, Eimas & Tarr, 2001). What these studies can’t tell us is what underlying neural processes are driving the emergence of such behaviours. For this, neuroimaging can complement the progress made in developmental psychology. It enables us to answer questions about when and where in the brain a capacity to distinguish a category is present and to chart this emergence in different brain regions. In adults, there are dedicated areas on the bottom – or ventral – surface that demonstrate reliable patterns of responses to the different objects we see in the world (Grill-Spector & Weiner, 2014). This series of brain regions is organised as a hierarchy, processing increasingly complex visual features, from low-level perceptual features such as orientation (Hubel & Wiesel, 1962) to more abstract groupings by semantic concept (Huth et al., 2012) or animacy (Konkle & Caramazza, 2013). Although infants’ looking preferences have been shown to align with these organising principles (Spriet et al., 2022), it remained unclear whether activity in these brain areas is present from an early age.

Beyond knowing when something becomes functional in the infant brain, we also would like to know exactly how this comes to be. What learning mechanisms are at play? Is it the kind of structured information that can be learned from the statistics of the environment, or is a strong innate prior required? For this, a computational model of learning would be useful. Indeed, a further benefit of studying the development of the visual system, in addition to the wealth of research on infant looking behaviour to which we can compare, is the existence of advanced computational models of vision. For example, we can use deep neural network models that learn how to recognise objects through a training process during which they see many images. After training, the model’s hierarchical layers map onto the ventral visual hierarchy, processing increasingly complex visual features similarly to the brain (Yamins et al., 2014). We wondered whether such models would also capture visual responses in the infant brain. If so, we could test if the infant brain might map onto models at an earlier stage of their training, or perhaps earlier layers of the model that process simpler features. We believed that if such models could effectively capture infant brain responses, this would open new avenues for building advanced computational models of infant brain development. Models whose learning can be tightly controlled may provide insights into the unique processing of the infant brain that we cannot determine from comparison to adults alone.

To answer these questions we needed rich, visual responses from the infant brain that are suitable for comparison with computational models. We therefore required data from a large cohort of awake infants as they viewed a variety of visual categories. Collecting high-quality neuroimaging data in infants comes with many challenges. MRI requires participants to refrain from moving, but we cannot ask babies to lie still. Participants may also become fussy if a particular task is not interesting, or they may fall asleep mid experiment. Despite these challenges, pioneering work in awake infant fMRI laid the foundations for what was possible in this space (Deen et al., 2017, Ellis et al., 2020). These studies provided evidence that category selectivity to faces, scenes and body parts is present from early infancy (Kosakowski et al., 2022), as is the basic structure of early visual regions that reflect where in our visual field we perceive an object (Ellis et al., 2021). Given this, we sought to characterise infants’ responses to a broader set of objects, including animate and inanimate categories such as cats, rubber ducks, and trees. This diverse set of objects allowed us to measure the fine-grained patterns of response required for comparison with deep neural network models.
For our study, 134 families generously took the time to participate when their newest family member was just 2 months of age (see Figure 1). We successfully collected awake fMRI data from 130 of these infants as they watched our visual stimuli in the scanner. Encouragingly, many infants engaged with the images depicting various objects. After careful data preprocessing and quality control for motion we had a final dataset of 101 2 month olds who completed this experiment. The same infants were also scanned at 9 months, which gave us a final sample of 44 at this age group. This dataset allowed us to robustly characterise developing infant brain responses in ways that were not previously possible. Using patterns of activity across a given brain region, we found that category structure is present by 2-months of age. Although early visual regions were the most mature, infants’ high-level visual regions were organised very similarly to adults. There was also distinct processing for categories that was reliable across different examples of the same object. Fascinatingly, the infants’ brain responses already aligned with the deepest layers of a neural network, indicating complex processing beyond simple visual features. This suggests that the structure that’s present at this age is as complex as that learned by a model that has seen millions of images and can accurately label many objects.

Figure 1. 2-month-old participant with their mother participating in the study.

In addition to revealing that category structure is present in the youngest minds, much earlier than previously appreciated, this work proved that we can compare advanced computational models to developing brains. This is very exciting, as it unlocks a new toolkit that developmental scientists can use to reveal the mechanisms of information processing and learning during infancy, laying the groundwork for the emerging field of developmental cognitive computational neuroscience. Deep neural network models provide a computational language for asking questions such as: what rules does the infant brain use to learn new information? What information can be learned from experience alone? Which properties of the developing brain rely on innate structure or genetically predetermined wiring? Deep neural networks allow us to explore these questions by varying the model architecture, the data on which we train, and the rules that the models use to learn. By combining behaviour, neuroimaging, and computational modelling to provide a holistic account of development, this brings us one step closer to answering the question: what is going on inside an infant’s mind?

Deen, B., Richardson, H., Dilks, D. D., Takahashi, A., Keil, B., Wald, L. L., … & Saxe, R. (2017). Organization of high-level visual cortex in human infants. Nature communications, 8(1), 13995.
Ellis, C. T., Skalaban, L. J., Yates, T. S., Bejjanki, V. R., Córdova, N. I., & Turk-Browne, N. B. (2020). Re-imagining fMRI for awake behaving infants. Nature Communications, 11(1), 4523.
Ellis, C. T., Yates, T. S., Skalaban, L. J., Bejjanki, V. R., Arcaro, M. J., & Turk-Browne, N. B. (2021). Retinotopic organization of visual cortex in human infants. Neuron, 109(16), 2616-2626.
Grill-Spector, K., & Weiner, K. S. (2014). The functional architecture of the ventral temporal cortex and its role in categorization. Nature Reviews Neuroscience, 15(8), 536-548.
Hubel, D. H., & Wiesel, T. N. (1962). Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex. The Journal of Physiology, 160(1), 106–154.
Huth, A. G., Nishimoto, S., Vu, A. T., & Gallant, J. L. (2012). A continuous semantic space describes the representation of thousands of object and action categories across the human brain. Neuron, 76(6), 1210-1224.
Konkle, T., & Caramazza, A. (2013). Tripartite organization of the ventral stream by animacy and object size. Journal of Neuroscience, 33(25), 10235-10242.
Kosakowski, H. L., Cohen, M. A., Takahashi, A., Keil, B., Kanwisher, N., & Saxe, R. (2022). Selective responses to faces, scenes, and bodies in the ventral visual pathway of infants. Current Biology, 32(2), 265-274.
O’Doherty, C., Dineen, Á. T., Truzzi, A., King, G., Zaadnoordijk, L., Harrison, K., … & Cusack, R. (2026). Infants have rich visual categories in ventrotemporal cortex at 2 months of age. Nature Neuroscience, 1-10.
Quinn, P. C., Eimas, P. D., & Tarr, M. J. (2001). Perceptual categorization of cat and dog silhouettes by 3-to 4-month-old infants. Journal of experimental child psychology, 79(1), 78-94.
Spriet, C., Abassi, E., Hochmann, J. R., & Papeo, L. (2022). Visual object categorization in infancy. Proceedings of the National Academy of Sciences, 119(8), e2105866119.
Yamins, D. L., Hong, H., Cadieu, C. F., Solomon, E. A., Seibert, D., & DiCarlo, J. J. (2014). Performance-optimized hierarchical models predict neural responses in higher visual cortex. Proceedings of the national academy of sciences, 111(23), 8619-8624.

About the Author