Why models
It is hard to imagine what an infant's mind is like, and neuroimaging can only tell us so much on its own. Our approach is to build learning systems that face the same problem an infant does, namely making sense of an unlabelled stream of visual experience, and to ask whether what they learn matches what we see in the infant brain. Where a model succeeds, it gives us a candidate mechanism. Where it fails, it tells us what the infant has that the model lacks. The traffic runs both ways: infants learn with far less data and energy than current AI, and understanding how could lead to better learning algorithms.
InfantNeuroAI (2026–)
Our new programme, funded by an ERC Advanced Grant [101265741], has three goals. To understand how infants acquire the brain representations that form the foundations of cognition, across the whole first year, using awake fMRI and OPM-MEG. To develop neuroimaging methods that can detect disrupted function after neonatal brain injury. And to identify more energy-efficient learning algorithms that could reduce the economic and environmental cost of AI. The team combines neuroscientists, neuroimagers and AI engineers, and everyone works end to end, from designing experiments to building models.
What FOUNDCOG showed (2019–2025)
The programme grew out of the ERC-funded Foundations of Cognition project. Its central result: self-supervised networks trained on video from an infant's point of view predict infant visual cortex better than networks trained on adult-style datasets (O'Doherty et al., 2025, Cognitive Computational Neuroscience), and the categories such models learn line up with those we find in two-month-old ventrotemporal cortex (O'Doherty et al., 2026, Nature Neuroscience). FOUNDCOG also established the awake-infant paradigms now used across the lab, described in Dineen et al. (2026).
Current modelling work
Self-supervised learning from infant-view video, and how choice of objective and architecture changes what is learned. Continual reinforcement learning: how an agent can keep learning without forgetting, as infants must (Kim Nolle; Noelle et al., 2025). Developmentally inspired networks that acquire visual experience in the order infants do, and comparison of their internal representations with infant fMRI (Áine Dineen).
New cluster
This year, we installed a new GPU cluster comprising 24 H200 GPUs and 2 petabytes of secure storage, for brain health work.
Code and data
Our stack is Python throughout (PyTorch, NiPype, BIDS). The FOUNDCOG preprocessing and GLM pipelines are public at github.com/ClionaOD/foundcog_analysis; more at github.com/cusacklab.
Join us
We are recruiting two postdoctoral fellows and three PhD students to InfantNeuroAI; see vacancies.
It is hard to imagine what an infant's mind is like, and neuroimaging can only tell us so much on its own. Our approach is to build learning systems that face the same problem an infant does, namely making sense of an unlabelled stream of visual experience, and to ask whether what they learn matches what we see in the infant brain. Where a model succeeds, it gives us a candidate mechanism. Where it fails, it tells us what the infant has that the model lacks. The traffic runs both ways: infants learn with far less data and energy than current AI, and understanding how could lead to better learning algorithms.
InfantNeuroAI (2026–)
Our new programme, funded by an ERC Advanced Grant [101265741], has three goals. To understand how infants acquire the brain representations that form the foundations of cognition, across the whole first year, using awake fMRI and OPM-MEG. To develop neuroimaging methods that can detect disrupted function after neonatal brain injury. And to identify more energy-efficient learning algorithms that could reduce the economic and environmental cost of AI. The team combines neuroscientists, neuroimagers and AI engineers, and everyone works end to end, from designing experiments to building models.
What FOUNDCOG showed (2019–2025)
The programme grew out of the ERC-funded Foundations of Cognition project. Its central result: self-supervised networks trained on video from an infant's point of view predict infant visual cortex better than networks trained on adult-style datasets (O'Doherty et al., 2025, Cognitive Computational Neuroscience), and the categories such models learn line up with those we find in two-month-old ventrotemporal cortex (O'Doherty et al., 2026, Nature Neuroscience). FOUNDCOG also established the awake-infant paradigms now used across the lab, described in Dineen et al. (2026).
Current modelling work
Self-supervised learning from infant-view video, and how choice of objective and architecture changes what is learned. Continual reinforcement learning: how an agent can keep learning without forgetting, as infants must (Kim Nolle; Noelle et al., 2025). Developmentally inspired networks that acquire visual experience in the order infants do, and comparison of their internal representations with infant fMRI (Áine Dineen).
New cluster
This year, we installed a new GPU cluster comprising 24 H200 GPUs and 2 petabytes of secure storage, for brain health work.
Code and data
Our stack is Python throughout (PyTorch, NiPype, BIDS). The FOUNDCOG preprocessing and GLM pipelines are public at github.com/ClionaOD/foundcog_analysis; more at github.com/cusacklab.
Join us
We are recruiting two postdoctoral fellows and three PhD students to InfantNeuroAI; see vacancies.
