Periodic Reporting for period 5 - FOUNDCOG (Curiosity and the Development of the Hidden Foundations of Cognition)
Berichtszeitraum: 2025-01-01 bis 2025-06-30
Human infants initially develop slowly, compared with many other animals. We hypothesize that this slow early development lays the foundations for cognition, and is critical to the flexibility and generalisation that characterise human intelligence.
Our project will develop computational models of infants’ acquisition of early knowledge, using deep neural networks, which currently dominate machine learning. We will then test the predictions of the models, using neuroimaging of human infants with magnetic resonance imaging (MRI).
The goals are to understand the development of the human mind in healthy infants, and how it is disrupted by brain injury. We will recruit a cohort of healthy infants from the maternity ward and measure development longitudinally, at 2 and 9 months of age, using MRI and online testing. We will also recruit a second cohort of infants from the neonatal intensive care unit, who are at an elevated risk of developing cognitive, behavioural and social impairments later in life, and contrast their brain development at the same points during the first year.
The overall objectives are to develop the scientific understanding of development in the helpless period of infancy and to understand how this may be disrupted. This may lead to future possible interventions to reduce the risk of developmental impairments. Furthermore, understanding how human infants learn should inspire new directions for machine learning.
A summary of the major achievements of the project:
(1) Acquire the largest longitudinal cohort of awake infants during fMRI to date (N=134).
(2) A paper that for the first time uses multi-voxel pattern analysis to study infant visual development and compare it to deep neural networks (O'Doherty et al, in press, Nature Neuroscience)
(3) Many studies of various aspects of infants' brain function are close to submission.
(4) The dataset has created considerable excitement and three other groups that are collaboratively analysing our data. We expect many more as the data start to be published.
Our project will develop computational models of infants’ acquisition of early knowledge, using deep neural networks, which currently dominate machine learning. We will then test the predictions of the models, using neuroimaging of human infants with magnetic resonance imaging (MRI).
The goals are to understand the development of the human mind in healthy infants, and how it is disrupted by brain injury. We will recruit a cohort of healthy infants from the maternity ward and measure development longitudinally, at 2 and 9 months of age, using MRI and online testing. We will also recruit a second cohort of infants from the neonatal intensive care unit, who are at an elevated risk of developing cognitive, behavioural and social impairments later in life, and contrast their brain development at the same points during the first year.
The overall objectives are to develop the scientific understanding of development in the helpless period of infancy and to understand how this may be disrupted. This may lead to future possible interventions to reduce the risk of developmental impairments. Furthermore, understanding how human infants learn should inspire new directions for machine learning.
A summary of the major achievements of the project:
(1) Acquire the largest longitudinal cohort of awake infants during fMRI to date (N=134).
(2) A paper that for the first time uses multi-voxel pattern analysis to study infant visual development and compare it to deep neural networks (O'Doherty et al, in press, Nature Neuroscience)
(3) Many studies of various aspects of infants' brain function are close to submission.
(4) The dataset has created considerable excitement and three other groups that are collaboratively analysing our data. We expect many more as the data start to be published.
*Self-supervised Babies
Developmental psychologists have shown that infants are learning many things in their first year. However, linguistic understanding is primitive until the end of the year, and so their learning must be "self-supervised", in that they can learn without being explicitly taught. At present, machines are mostly taught using hand-curated datasets, which are painstakingly labelled by humans. Self-supervised learning algorithms can potentially reduce the dependence on these datasets, and so are of great interest to the machine learning community. With a machine learning expert at Google Deepmind and an evolution-development expert at Auburn University, we published a paper in "Trends in Cognitive Sciences" which has generated considerable interest (Cusack, Ranzato and Charvet, 2024), incluing a commentary and a reply from us. In a Nature Machine Intelligence article, Lorijn Zaadnoordijk from the lab, and our collaborator Tarek Besold have reviewed the developmental psychology literature to identify potential "next big thing(s)" for this area of machine learning (Zaadnoordijk, Besold & Cusack, 2022).
*Learning Visual Cateogries
Humans have a deep understanding of the world. When we recognise an object, we know what other things it is similar to and we can classify it as part of some superordinate category. This type of knowledge is called semantic knowledge. In a Nature Neuroscience paper (O'Doherty et al, in press), a student presents data showing unexpected maturity of visual representations at two months. Further, Cliona O'Doherty has been testing the idea that by observing the co-occurrences of objects in the world, infants could not just learn how to recognise things, but also learn about semantics. She has done this by setting up a computational model using a deep neural network. Cliona O'Doherty presented SemanticCMC - improved semantic self-supervised learning with naturalistic temporal co-occurrences at the workshop Self-supervised learning: theory and practice at Neural Information Processing Systems (NeurIPS) 2020.
*Timescales in the infant brain
In adults, brain regions have characteristic "intrinsic timescales" that reflect how long they retain information. Sensory regions tend to have short timescales, allowing them to rapidly update as the world changes. Higher brain regions have slower timescales, allowing them to hold on to memories for longer periods. We have studied these timescales using fMRI data in infants, and found that infants have overall slower timescales, and a different pattern across the brain. We have then collaborated with a group in Spain to extend these results to another imaging modality (EEG).
*How Can Random Networks Explain the Brain So Well?
A part of the brain called the inferotemporal (IT) cortex is critical for humans and other monkeys to visually recognise objects. Currently, deep neural networks are the best models of brain responses in the IT cortex of adults. It has been argued that this is because the visual features that deep neural networks learn for object recognition are the same as those IT uses. But, Anna Truzzi has been investigating a conundrum, which is that actually untrained (or random) deep neural networks also do a surprisingly good job of modelling IT activity.
Anna presented the paper "Convolutional Neural Networks as a Model of Visual Activity in The Brain: Greater Contribution of Architecture Than Learned Weights" at the workshop Bridging AI and Cognitive Science at the International Conference on Learning Representations (ICLR) 2020. She will also be presenting at the NeurIPS2020 workshop Shared Visual Representations in Humans and Machine Intelligence, with the title "Understanding CNNs as a model of the inferior temporal cortex: using mediation analysis to unpack the contribution of perceptual and semantic features in random and trained networks". This work is also directly relevant to neuroscientists, and was presented at the neuromatch 1.0 conference with the title "Are deep neural networks effective models of visual activity in the brain because of their architecture or training?".
Developmental psychologists have shown that infants are learning many things in their first year. However, linguistic understanding is primitive until the end of the year, and so their learning must be "self-supervised", in that they can learn without being explicitly taught. At present, machines are mostly taught using hand-curated datasets, which are painstakingly labelled by humans. Self-supervised learning algorithms can potentially reduce the dependence on these datasets, and so are of great interest to the machine learning community. With a machine learning expert at Google Deepmind and an evolution-development expert at Auburn University, we published a paper in "Trends in Cognitive Sciences" which has generated considerable interest (Cusack, Ranzato and Charvet, 2024), incluing a commentary and a reply from us. In a Nature Machine Intelligence article, Lorijn Zaadnoordijk from the lab, and our collaborator Tarek Besold have reviewed the developmental psychology literature to identify potential "next big thing(s)" for this area of machine learning (Zaadnoordijk, Besold & Cusack, 2022).
*Learning Visual Cateogries
Humans have a deep understanding of the world. When we recognise an object, we know what other things it is similar to and we can classify it as part of some superordinate category. This type of knowledge is called semantic knowledge. In a Nature Neuroscience paper (O'Doherty et al, in press), a student presents data showing unexpected maturity of visual representations at two months. Further, Cliona O'Doherty has been testing the idea that by observing the co-occurrences of objects in the world, infants could not just learn how to recognise things, but also learn about semantics. She has done this by setting up a computational model using a deep neural network. Cliona O'Doherty presented SemanticCMC - improved semantic self-supervised learning with naturalistic temporal co-occurrences at the workshop Self-supervised learning: theory and practice at Neural Information Processing Systems (NeurIPS) 2020.
*Timescales in the infant brain
In adults, brain regions have characteristic "intrinsic timescales" that reflect how long they retain information. Sensory regions tend to have short timescales, allowing them to rapidly update as the world changes. Higher brain regions have slower timescales, allowing them to hold on to memories for longer periods. We have studied these timescales using fMRI data in infants, and found that infants have overall slower timescales, and a different pattern across the brain. We have then collaborated with a group in Spain to extend these results to another imaging modality (EEG).
*How Can Random Networks Explain the Brain So Well?
A part of the brain called the inferotemporal (IT) cortex is critical for humans and other monkeys to visually recognise objects. Currently, deep neural networks are the best models of brain responses in the IT cortex of adults. It has been argued that this is because the visual features that deep neural networks learn for object recognition are the same as those IT uses. But, Anna Truzzi has been investigating a conundrum, which is that actually untrained (or random) deep neural networks also do a surprisingly good job of modelling IT activity.
Anna presented the paper "Convolutional Neural Networks as a Model of Visual Activity in The Brain: Greater Contribution of Architecture Than Learned Weights" at the workshop Bridging AI and Cognitive Science at the International Conference on Learning Representations (ICLR) 2020. She will also be presenting at the NeurIPS2020 workshop Shared Visual Representations in Humans and Machine Intelligence, with the title "Understanding CNNs as a model of the inferior temporal cortex: using mediation analysis to unpack the contribution of perceptual and semantic features in random and trained networks". This work is also directly relevant to neuroscientists, and was presented at the neuromatch 1.0 conference with the title "Are deep neural networks effective models of visual activity in the brain because of their architecture or training?".
* The success rate of 97% in infants at 2 months was much higher than foreseen from the previous literature (~50%). This is likely due to both our innvoations in infant comfort, and to the engagement of infants at this age with our visual stimuli. Two methodological papers have been drafted and will soon be submitted.
* We have shown that awake infant fMRI and multi-voxel pattern analysis provides a rich new way to probe brain function in infants.
* We have found visual, memory and frontal systems are more mature than expected at 2 months.
* Deep neural networks have proven highly effective as computational models of infant brain function, and allow many new questions to be addressed.
* Our computational modelling aims to yield new directions in machine learning.
* We have shown that awake infant fMRI and multi-voxel pattern analysis provides a rich new way to probe brain function in infants.
* We have found visual, memory and frontal systems are more mature than expected at 2 months.
* Deep neural networks have proven highly effective as computational models of infant brain function, and allow many new questions to be addressed.
* Our computational modelling aims to yield new directions in machine learning.