Fosca Al Roumi
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Fosca Al Roumi has a total of 26 co-authors. Here are the top co-authors by number of shared publications:
- Lucas Benjamin
Laboratoire de Neurosciences Cognitives et Computationnelles, Inserm U 960, Paris, France
8 shared publications - Stanislas Dehaene
Cognitive Neuroimaging Unit, CEA, INSERM, Université Paris-Saclay, NeuroSpin center, 91191 Gif/Yvette, France
7 shared publications - Ghislaine Dehaene-Lambertz
Cognitive Neuroimaging Unit, CNRS ERL 9003, INSERM U992, CEA, Université Paris-Saclay, NeuroSpin center
4 shared publications Liping Wang
Institute of Neuroscience, Key Laboratory of Primate Neurobiology, CAS Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of Sciences
4 shared publications- Mathias Sablé-Meyer
Sainsbury Wellcome Centre, University College London
4 shared publications Chenxi He
Cognitive Neuroimaging Unit, CEA, INSERM, Université Paris-Saclay, NeuroSpin center
3 shared publications
Latest Publications
- Open Access
- 28 July 2026
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 123 (31)
Long-horizon associative learning as a unifying framework for statistical learning across scales
Sensory inputs are rich with temporal patterns that unfold across multiple timescales. Uncovering these regularities is essential for anticipating future events and navigating the environment efficiently. Numerous models have been proposed to account for learning at specific temporal scales; however, they are often designed in isolation and rely on narrowly tuned statistical measures, limiting their generalizability to other paradigms. In contrast, humans typically learn without prior knowledge of the underlying structure or the relevant timescale at which regularities occur. Here, we present a unifying account of statistical learning that spans a wide range of temporal dependencies, from adjacent and nonadjacent transitions to complex network structures. This model, long-horizon associative learning, offers a biologically grounded implementation of the successor representation, or equivalently, the free energy minimization model. Reanalyzing data from 11 previously published studies, we show that a single neural mechanism captures both local statistical regularities and higher-order structural properties. This mechanism rests on graded temporal overlap of associative traces and is governed by a single free parameter (β). This initial domain-general associative learning process, emerging from the graded structure of associations, may later scaffold to higher-level operations such as grouping, categorization, rule abstraction, and memory formation. Overall, this framework offers a conceptual synthesis that bridges disparate strands of the statistical learning literature and reframes apparent paradigm-specific effects as different expressions of a common underlying computation.
- 28 July 2026
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 123 (31)
Open AccessLong-horizon associative learning as a unifying framework for statistical learning across scales
Sensory inputs are rich with temporal patterns that unfold across multiple timescales. Uncovering these regularities is essential for anticipating future events and navigating the environment efficiently. Numerous models have been proposed to account for learning at specific temporal scales; however, they are often designed in isolation and rely on narrowly tuned statistical measures, limiting their generalizability to other paradigms. In contrast, humans typically learn without prior knowledge of the underlying structure or the relevant timescale at which regularities occur. Here, we present a unifying account of statistical learning that spans a wide range of temporal dependencies, from adjacent and nonadjacent transitions to complex network structures. This model, long-horizon associative learning, offers a biologically grounded implementation of the successor representation, or equivalently, the free energy minimization model. Reanalyzing data from 11 previously published studies, we show that a single neural mechanism captures both local statistical regularities and higher-order structural properties. This mechanism rests on graded temporal overlap of associative traces and is governed by a single free parameter (β). This initial domain-general associative learning process, emerging from the graded structure of associations, may later scaffold to higher-level operations such as grouping, categorization, rule abstraction, and memory formation. Overall, this framework offers a conceptual synthesis that bridges disparate strands of the statistical learning literature and reframes apparent paradigm-specific effects as different expressions of a common underlying computation.
- Open Access
An ordinal Language of Thought supports human memory for regular sequences
How do humans store sequences that far exceed working memory capacity? Using visuo-spatial and binary auditory sequences, we previously showed that a Language of Thought (LoT) architecture, in which simple primitives are recursively combined into hierarchical programs, enables efficient storage of structured sequences. Here we ask whether this principle extends to purely ordinal structure: sequences defined by how items repeat and in what order, as in AABBCCAABBCC, independently of their spatial content. Across three experiments, participants reproduced 12-item sequences of spatial locations with various ordinal structures. The minimal description length derived from the LoT model predicted recall accuracy with remarkable precision (r = .96), substantially outperforming Shannon entropy, Lempel-Ziv complexity, chunking models and subjective complexity ratings. Critically, fine-grained analyses of participants' inter-click intervals during reproduction revealed systematic slowdowns at the hierarchical boundaries predicted by the LoT programs, providing a behavioral signature of the underlying mental syntax. These results identify a compact vocabulary of mental primitives, repetition, mirroring, and interleaving, whose composition accounts for the symbolic compression of ordinal structures. For ordinal regularities, human sequence memory operates as a form of program induction, leveraging a domain-general capacity for hierarchical compression to encode complex structured information.
Open AccessAn ordinal Language of Thought supports human memory for regular sequences
How do humans store sequences that far exceed working memory capacity? Using visuo-spatial and binary auditory sequences, we previously showed that a Language of Thought (LoT) architecture, in which simple primitives are recursively combined into hierarchical programs, enables efficient storage of structured sequences. Here we ask whether this principle extends to purely ordinal structure: sequences defined by how items repeat and in what order, as in AABBCCAABBCC, independently of their spatial content. Across three experiments, participants reproduced 12-item sequences of spatial locations with various ordinal structures. The minimal description length derived from the LoT model predicted recall accuracy with remarkable precision (r = .96), substantially outperforming Shannon entropy, Lempel-Ziv complexity, chunking models and subjective complexity ratings. Critically, fine-grained analyses of participants' inter-click intervals during reproduction revealed systematic slowdowns at the hierarchical boundaries predicted by the LoT programs, providing a behavioral signature of the underlying mental syntax. These results identify a compact vocabulary of mental primitives, repetition, mirroring, and interleaving, whose composition accounts for the symbolic compression of ordinal structures. For ordinal regularities, human sequence memory operates as a form of program induction, leveraging a domain-general capacity for hierarchical compression to encode complex structured information.
- Open Access
- 12 January 2026
- Published by eLife Sciences Publications, Ltd in eLife
- Vol. 14
A geometric shape regularity effect in the human brain
The perception and production of regular geometric shapes, a characteristic trait of human cultures since prehistory, has unknown neural mechanisms. Behavioral studies suggest that humans are attuned to discrete regularities such as symmetries and parallelism and rely on their combinations to encode regular geometric shapes in a compressed form. To identify the brain systems underlying this ability, as well as their dynamics, we collected functional MRI in both adults and 6-year-olds, and magnetoencephalography data in adults, during the perception of simple shapes such as hexagons, triangles, and quadrilaterals. The results revealed that geometric shapes, relative to other visual categories, induce a hypoactivation of ventral visual areas and an overactivation of the intraparietal and inferior temporal regions also involved in mathematical processing, whose activation is modulated by geometric regularity. While convolutional neural networks captured the early visual activity evoked by geometric shapes, they failed to account for subsequent dorsal parietal and prefrontal signals, which could only be captured by discrete geometric features or by bigger deep-learning models of vision. We propose that the perception of abstract geometric regularities engages an additional symbolic mode of visual perception.
- 12 January 2026
- Published by eLife Sciences Publications, Ltd in eLife
- Vol. 14
Open AccessA geometric shape regularity effect in the human brain
The perception and production of regular geometric shapes, a characteristic trait of human cultures since prehistory, has unknown neural mechanisms. Behavioral studies suggest that humans are attuned to discrete regularities such as symmetries and parallelism and rely on their combinations to encode regular geometric shapes in a compressed form. To identify the brain systems underlying this ability, as well as their dynamics, we collected functional MRI in both adults and 6-year-olds, and magnetoencephalography data in adults, during the perception of simple shapes such as hexagons, triangles, and quadrilaterals. The results revealed that geometric shapes, relative to other visual categories, induce a hypoactivation of ventral visual areas and an overactivation of the intraparietal and inferior temporal regions also involved in mathematical processing, whose activation is modulated by geometric regularity. While convolutional neural networks captured the early visual activity evoked by geometric shapes, they failed to account for subsequent dorsal parietal and prefrontal signals, which could only be captured by discrete geometric features or by bigger deep-learning models of vision. We propose that the perception of abstract geometric regularities engages an additional symbolic mode of visual perception.
- Open Access
- 26 November 2025
- Published by eLife Sciences Publications, Ltd in eLife preprints
A geometric shape regularity effect in the human brain
The perception and production of regular geometric shapes, a characteristic trait of human cultures since prehistory, has unknown neural mechanisms. Behavioral studies suggest that humans are attuned to discrete regularities such as symmetries and parallelism, and rely on their combinations to encode regular geometric shapes in a compressed form. To identify the brain systems underlying this ability, as well as their dynamics, we collected functional MRI in both adults and six-year-olds, and magnetoencephalography data in adults, during the perception of simple shapes such as hexagons, triangles and quadrilaterals. The results revealed that geometric shapes, relative to other visual categories, induce a hypoactivation of ventral visual areas and an overactivation of the intraparietal and inferior temporal regions also involved in mathematical processing, whose activation is modulated by geometric regularity. While convolutional neural networks captured the early visual activity evoked by geometric shapes, they failed to account for subsequent dorsal parietal and prefrontal signals, which could only be captured by discrete geometric features or by bigger deep-learning models of vision. We propose that the perception of abstract geometric regularities engages an additional symbolic mode of visual perception.
- 26 November 2025
- Published by eLife Sciences Publications, Ltd in eLife preprints
Open AccessA geometric shape regularity effect in the human brain
The perception and production of regular geometric shapes, a characteristic trait of human cultures since prehistory, has unknown neural mechanisms. Behavioral studies suggest that humans are attuned to discrete regularities such as symmetries and parallelism, and rely on their combinations to encode regular geometric shapes in a compressed form. To identify the brain systems underlying this ability, as well as their dynamics, we collected functional MRI in both adults and six-year-olds, and magnetoencephalography data in adults, during the perception of simple shapes such as hexagons, triangles and quadrilaterals. The results revealed that geometric shapes, relative to other visual categories, induce a hypoactivation of ventral visual areas and an overactivation of the intraparietal and inferior temporal regions also involved in mathematical processing, whose activation is modulated by geometric regularity. While convolutional neural networks captured the early visual activity evoked by geometric shapes, they failed to account for subsequent dorsal parietal and prefrontal signals, which could only be captured by discrete geometric features or by bigger deep-learning models of vision. We propose that the perception of abstract geometric regularities engages an additional symbolic mode of visual perception.
- Open Access
Long-Horizon associative learning as a unifying framework for statistical learning across scales
Sensory inputs are rich with temporal patterns that unfold across multiple timescales. Discovering these regularities is essential for predicting future events and navigating the environment efficiently. Numerous models have been proposed to account for learning at specific temporal scales; however, they are often designed in isolation and rely on narrowly tuned statistical measures, limiting their generalizability to other paradigms. In contrast, humans typically learn without prior knowledge of the underlying structure or the relevant timescale at which regularities occur. Here, we present a unifying account of statistical learning that spans a wide range of temporal dependencies, from adjacent and non-adjacent transitions to complex network structures. This model, Long-Horizon Associative Learning (L-HAL), offers a biologically grounded implementation of the successor representation, or equivalently, the Free Energy Minimization Model. Reanalyzing data from 11 previously published studies, we show that a single neural mechanism, based on graded temporal overlap of associative traces and governed by a single free parameter (β), captures both local statistical regularities and higher-order structural properties. This initial domain-general associative learning process, emerging from the graded structure of associations, may later scaffold to higher-level operations such as grouping, categorization, rule abstraction, and memory formation. Overall, this framework offers a conceptual synthesis that bridges disparate strands of the statistical learning literature and reframes apparent paradigm-specific effects as different expressions of a common underlying computation
Open AccessLong-Horizon associative learning as a unifying framework for statistical learning across scales
Sensory inputs are rich with temporal patterns that unfold across multiple timescales. Discovering these regularities is essential for predicting future events and navigating the environment efficiently. Numerous models have been proposed to account for learning at specific temporal scales; however, they are often designed in isolation and rely on narrowly tuned statistical measures, limiting their generalizability to other paradigms. In contrast, humans typically learn without prior knowledge of the underlying structure or the relevant timescale at which regularities occur. Here, we present a unifying account of statistical learning that spans a wide range of temporal dependencies, from adjacent and non-adjacent transitions to complex network structures. This model, Long-Horizon Associative Learning (L-HAL), offers a biologically grounded implementation of the successor representation, or equivalently, the Free Energy Minimization Model. Reanalyzing data from 11 previously published studies, we show that a single neural mechanism, based on graded temporal overlap of associative traces and governed by a single free parameter (β), captures both local statistical regularities and higher-order structural properties. This initial domain-general associative learning process, emerging from the graded structure of associations, may later scaffold to higher-level operations such as grouping, categorization, rule abstraction, and memory formation. Overall, this framework offers a conceptual synthesis that bridges disparate strands of the statistical learning literature and reframes apparent paradigm-specific effects as different expressions of a common underlying computation
- Open Access
A geometric shape regularity effect in the human brain
The perception and production of regular geometric shapes, a characteristic trait of human cultures since prehistory, has unknown neural mechanisms. Behavioral studies suggest that humans are attuned to discrete regularities such as symmetries and parallelism, and rely on their combinations to encode regular geometric shapes in a compressed form. To identify the brain systems underlying this ability, as well as their dynamics, we collected functional MRI in both adults and six-year-olds, and magnetoencephalography data in adults, during the perception of simple shapes such as hexagons, triangles and quadrilaterals. The results revealed that geometric shapes, relative to other visual categories, induce a hypoactivation of ventral visual areas and an overactivation of the intraparietal and inferior temporal regions also involved in mathematical processing, whose activation is modulated by geometric regularity. While convolutional neural networks captured the early visual activity evoked by geometric shapes, they failed to account for subsequent dorsal parietal and prefrontal signals, which could only be captured by discrete geometric features or by bigger deep-learning models of vision. We propose that the perception of abstract geometric regularities engages an additional symbolic mode of visual perception.
Open AccessA geometric shape regularity effect in the human brain
The perception and production of regular geometric shapes, a characteristic trait of human cultures since prehistory, has unknown neural mechanisms. Behavioral studies suggest that humans are attuned to discrete regularities such as symmetries and parallelism, and rely on their combinations to encode regular geometric shapes in a compressed form. To identify the brain systems underlying this ability, as well as their dynamics, we collected functional MRI in both adults and six-year-olds, and magnetoencephalography data in adults, during the perception of simple shapes such as hexagons, triangles and quadrilaterals. The results revealed that geometric shapes, relative to other visual categories, induce a hypoactivation of ventral visual areas and an overactivation of the intraparietal and inferior temporal regions also involved in mathematical processing, whose activation is modulated by geometric regularity. While convolutional neural networks captured the early visual activity evoked by geometric shapes, they failed to account for subsequent dorsal parietal and prefrontal signals, which could only be captured by discrete geometric features or by bigger deep-learning models of vision. We propose that the perception of abstract geometric regularities engages an additional symbolic mode of visual perception.
- Open Access
Long-horizon associative learning explains human sensitivity to statistical and network structures in auditory sequences
Networks are a useful mathematical tool for capturing the complexity of the world. In a previous behavioral study, we showed that human adults were sensitive to the high-level network structure underlying auditory sequences, even when presented with incomplete information. Their performance was best explained by a mathematical model compatible with associative learning principles, based on the integration of the transition probabilities between adjacent and non-adjacent elements with a memory decay. In the present study, we explored the neural correlates of this hypothesis via magnetoencephalography (MEG). Participants passively listened to sequences of tones organized in a sparse community network structure comprising two communities. An early difference (~150 ms) was observed in the brain responses to tone transitions with similar transition probability but occurring either within or between communities. This result implies a rapid and automatic encoding of the sequence structure. Using time-resolved decoding, we estimated the duration and overlap of the representation of each tone. The decoding performance exhibited exponential decay, resulting in a significant overlap between the representations of successive tones. Based on this extended decay profile, we estimated a long-horizon associative learning novelty index for each transition and found a correlation of this measure with the MEG signal. Overall, our study sheds light on the neural mechanisms underlying human sensitivity to network structures and highlights the potential role of Hebbian-like mechanisms in supporting learning at various temporal scales.
Open AccessLong-horizon associative learning explains human sensitivity to statistical and network structures in auditory sequences
Networks are a useful mathematical tool for capturing the complexity of the world. In a previous behavioral study, we showed that human adults were sensitive to the high-level network structure underlying auditory sequences, even when presented with incomplete information. Their performance was best explained by a mathematical model compatible with associative learning principles, based on the integration of the transition probabilities between adjacent and non-adjacent elements with a memory decay. In the present study, we explored the neural correlates of this hypothesis via magnetoencephalography (MEG). Participants passively listened to sequences of tones organized in a sparse community network structure comprising two communities. An early difference (~150 ms) was observed in the brain responses to tone transitions with similar transition probability but occurring either within or between communities. This result implies a rapid and automatic encoding of the sequence structure. Using time-resolved decoding, we estimated the duration and overlap of the representation of each tone. The decoding performance exhibited exponential decay, resulting in a significant overlap between the representations of successive tones. Based on this extended decay profile, we estimated a long-horizon associative learning novelty index for each transition and found a correlation of this measure with the MEG signal. Overall, our study sheds light on the neural mechanisms underlying human sensitivity to network structures and highlights the potential role of Hebbian-like mechanisms in supporting learning at various temporal scales.
- Top Cited PapersOpen Access
- 1 November 2023
- Published by eLife Sciences Publications, Ltd in eLife
- Vol. 12
Brain-imaging evidence for compression of binary sound sequences in human memory
According to the language-of-thought hypothesis, regular sequences are compressed in human memory using recursive loops akin to a mental program that predicts future items. We tested this theory by probing memory for 16-item sequences made of two sounds. We recorded brain activity with functional MRI and magneto-encephalography (MEG) while participants listened to a hierarchy of sequences of variable complexity, whose minimal description required transition probabilities, chunking, or nested structures. Occasional deviant sounds probed the participants’ knowledge of the sequence. We predicted that task difficulty and brain activity would be proportional to the complexity derived from the minimal description length in our formal language. Furthermore, activity should increase with complexity for learned sequences, and decrease with complexity for deviants. These predictions were upheld in both fMRI and MEG, indicating that sequence predictions are highly dependent on sequence structure and become weaker and delayed as complexity increases. The proposed language recruited bilateral superior temporal, precentral, anterior intraparietal, and cerebellar cortices. These regions overlapped extensively with a localizer for mathematical calculation, and much less with spoken or written language processing. We propose that these areas collectively encode regular sequences as repetitions with variations and their recursive composition into nested structures.
- 1 November 2023
- Published by eLife Sciences Publications, Ltd in eLife
- Vol. 12
Top Cited PapersOpen AccessBrain-imaging evidence for compression of binary sound sequences in human memory
According to the language-of-thought hypothesis, regular sequences are compressed in human memory using recursive loops akin to a mental program that predicts future items. We tested this theory by probing memory for 16-item sequences made of two sounds. We recorded brain activity with functional MRI and magneto-encephalography (MEG) while participants listened to a hierarchy of sequences of variable complexity, whose minimal description required transition probabilities, chunking, or nested structures. Occasional deviant sounds probed the participants’ knowledge of the sequence. We predicted that task difficulty and brain activity would be proportional to the complexity derived from the minimal description length in our formal language. Furthermore, activity should increase with complexity for learned sequences, and decrease with complexity for deviants. These predictions were upheld in both fMRI and MEG, indicating that sequence predictions are highly dependent on sequence structure and become weaker and delayed as complexity increases. The proposed language recruited bilateral superior temporal, precentral, anterior intraparietal, and cerebellar cortices. These regions overlapped extensively with a localizer for mathematical calculation, and much less with spoken or written language processing. We propose that these areas collectively encode regular sequences as repetitions with variations and their recursive composition into nested structures.
- Open Access
Nonperturbative collapse models for collisionless self-gravitating flows
Structure formation in the Universe has been well-studied within the Eulerian and Lagrangian perturbation theories, where the latter performs substantially better in comparison with N-body simulations. Standing out is the celebrated Zel'dovich approximation for dust matter. In this work, we recall the description of gravitational noncollisional systems and extend both the Eulerian and Lagrangian approaches by including, possibly anisotropic, velocity dispersion. A simple case with plane symmetry is then studied with an exact, nonperturbative approach, and various approximations of the derived model are then compared numerically. A striking result is that linearized Lagrangian solutions outperform models based on Burgers' equation in the multistream regime in comparison with the exact solution. These results are finally extended to a 3D case without symmetries, and master equations for the evolution of all parts of the perturbations are derived. The particular 3D case studied corresponds to a maximally anisotropic collapse, which involves an approximation based on the estimation of importance of the different levels of spatial derivatives of the local deformation field.
Open AccessNonperturbative collapse models for collisionless self-gravitating flows
Structure formation in the Universe has been well-studied within the Eulerian and Lagrangian perturbation theories, where the latter performs substantially better in comparison with N-body simulations. Standing out is the celebrated Zel'dovich approximation for dust matter. In this work, we recall the description of gravitational noncollisional systems and extend both the Eulerian and Lagrangian approaches by including, possibly anisotropic, velocity dispersion. A simple case with plane symmetry is then studied with an exact, nonperturbative approach, and various approximations of the derived model are then compared numerically. A striking result is that linearized Lagrangian solutions outperform models based on Burgers' equation in the multistream regime in comparison with the exact solution. These results are finally extended to a 3D case without symmetries, and master equations for the evolution of all parts of the perturbations are derived. The particular 3D case studied corresponds to a maximally anisotropic collapse, which involves an approximation based on the estimation of importance of the different levels of spatial derivatives of the local deformation field.
- 2 October 2023
- Published by American Physical Society (APS) in Physical Review D
- Vol. 108 (8) , 083502
Nonperturbative collapse models for collisionless self-gravitating flows
Structure formation in the Universe has been well studied within the Eulerian and Lagrangian perturbation theories, where the latter performs substantially better in comparison with N-body simulations. Standing out is the celebrated Zel’dovich approximation for dust matter. In this work, we recall the description of gravitational noncollisional systems and extend both the Eulerian and Lagrangian approaches by including, possibly anisotropic, velocity dispersion. A simple case with plane symmetry is then studied with an exact, nonperturbative approach, and various approximations of the derived model are then compared numerically. A striking result is that linearized Lagrangian solutions outperform models based on Burgers’ equation in the multistream regime in comparison with the exact solution. These results are finally extended to a 3D case without symmetries, and master equations for the evolution of all parts of the perturbations are derived. The particular 3D case studied corresponds to a maximally anisotropic collapse, which involves an approximation based on the estimation of importance of the different levels of spatial derivatives of the local deformation field.
- 2 October 2023
- Published by American Physical Society (APS) in Physical Review D
- Vol. 108 (8) , 083502
Nonperturbative collapse models for collisionless self-gravitating flows
Structure formation in the Universe has been well studied within the Eulerian and Lagrangian perturbation theories, where the latter performs substantially better in comparison with N-body simulations. Standing out is the celebrated Zel’dovich approximation for dust matter. In this work, we recall the description of gravitational noncollisional systems and extend both the Eulerian and Lagrangian approaches by including, possibly anisotropic, velocity dispersion. A simple case with plane symmetry is then studied with an exact, nonperturbative approach, and various approximations of the derived model are then compared numerically. A striking result is that linearized Lagrangian solutions outperform models based on Burgers’ equation in the multistream regime in comparison with the exact solution. These results are finally extended to a 3D case without symmetries, and master equations for the evolution of all parts of the perturbations are derived. The particular 3D case studied corresponds to a maximally anisotropic collapse, which involves an approximation based on the estimation of importance of the different levels of spatial derivatives of the local deformation field.