Reconstructing computational system dynamics from neural data with recurrent neural networks

Nature reviews. Neuroscience, Год журнала: 2023, Номер 24(11), С. 693 - 710

Опубликована: Окт. 4, 2023

Язык: Английский

---

Computational role of structure in neural activity and connectivity

Trends in Cognitive Sciences, Год журнала: 2024, Номер 28(7), С. 677 - 690

Опубликована: Март 28, 2024

Язык: Английский

---

Parallel movement planning is achieved via an optimal preparatory state in motor cortex

Cell Reports, Год журнала: 2023, Номер 42(2), С. 112136 - 112136

Опубликована: Фев. 1, 2023

How do patterns of neural activity in the motor cortex contribute to planning a movement? A recent theory developed for single movements proposes that acts as dynamical system whose initial state is optimized during preparatory phase movement. This makes important yet untested predictions about dynamics more complex behavioral settings. Here, we analyze non-human primates not one but two simultaneously. As predicted by theory, find parallel achieved adjusting within an optimal subspace intermediate reflecting trade-off between movements. The quantitatively accounts relationship this and fluctuations animals' behavior down at trial level. These results uncover simple mechanism multiple further point controlled process.

Язык: Английский

---

Emergent neural dynamics and geometry for generalization in a transitive inference task

PLoS Computational Biology, Год журнала: 2024, Номер 20(4), С. e1011954 - e1011954

Опубликована: Апрель 25, 2024

Relational cognition—the ability to infer relationships that generalize novel combinations of objects—is fundamental human and animal intelligence. Despite this importance, it remains unclear how relational cognition is implemented in the brain due part a lack hypotheses predictions at levels collective neural activity behavior. Here we discovered, analyzed, experimentally tested networks (NNs) perform transitive inference (TI), classic task (if A > B C, then C). We found NNs (i) generalized perfectly, despite lacking overt structure prior training, (ii) when required working memory (WM), capacity thought be essential brain, (iii) emergently expressed behaviors long observed living subjects, addition order-dependent behavior, (iv) different solutions yielding alternative behavioral predictions. Further, large-scale experiment, subjects performing WM-based TI showed behavior inconsistent with class characteristically an intuitive solution. These findings provide insights into classical ability, wider implications for realizes cognition.

Язык: Английский

---

Learning long-term motor timing/patterns on an orthogonal basis in random neural networks

Neural Networks, Год журнала: 2023, Номер 163, С. 298 - 311

Опубликована: Апрель 12, 2023

The ability of the brain to generate complex spatiotemporal patterns with specific timings is essential for motor learning and temporal processing. An approach that can model this function, using spontaneous activity a random neural network (RNN), associated orbital instability. We propose simple system learns an arbitrary time series as linear sum stable trajectories produced by several small modules. New finding in computer experiments module outputs are orthogonal each other. They created dynamic basis acquiring high representational capacity, which enabled learn timing extremely long intervals, such tens seconds millisecond computation unit, also Lorenz attractors. This self-sustained satisfies stability orthogonality requirements thus provides new neurocomputing framework perspective mechanisms learning.

Язык: Английский

---

MARBLE: interpretable representations of neural population dynamics using geometric deep learning

Nature Methods, Год журнала: 2025, Номер unknown

Опубликована: Фев. 17, 2025

Язык: Английский

---

Relating local connectivity and global dynamics in recurrent excitatory-inhibitory networks

PLoS Computational Biology, Год журнала: 2023, Номер 19(1), С. e1010855 - e1010855

Опубликована: Янв. 23, 2023

How the connectivity of cortical networks determines neural dynamics and resulting computations is one key questions in neuroscience. Previous works have pursued two complementary approaches to quantify structure connectivity. One approach starts from perspective biological experiments where only local statistics motifs between small groups neurons are accessible. Another based instead on artificial global matrix known, particular its low-rank can be used determine low-dimensional dynamics. A direct relationship these however currently missing. Specifically, it remains clarified how inter-related shape activity. To bridge this gap, here we develop a method for mapping onto an approximate structure. Our rests approximating using dominant eigenvectors, which compute perturbation theory random matrices. We demonstrate that multi-population defined central limit theorem holds approximated by with Gaussian-mixture statistics. specifically apply excitatory-inhibitory reciprocal motifs, show yields reliable predictions both dynamics, population Importantly, analytically accounts activity heterogeneity individual specific realizations Altogether, our allows us disentangle effects mean recurrent feedback, provides intuitive picture shapes network

Язык: Английский

---

Emergent perceptual biases from state-space geometry in trained spiking recurrent neural networks

Cell Reports, Год журнала: 2024, Номер 43(7), С. 114412 - 114412

Опубликована: Июль 1, 2024

A stimulus held in working memory is perceived as contracted toward the average stimulus. This contraction bias has been extensively studied psychophysics, but little known about its origin from neural activity. By training recurrent networks of spiking neurons to discriminate temporal intervals, we explored causes this and how behavior relates population firing We found that trained exhibited animal-like behavior. Various geometric features trajectories state space encoded warped representations durations first interval modulated by sensory history. Formulating a normative model, showed these conveyed Bayesian estimate durations, thus relating activity Importantly, our findings demonstrate computations already occur during phase persist throughout maintenance memory, until time comparison.

Язык: Английский

---

Dynamical mechanisms of how an RNN keeps a beat, uncovered with a low-dimensional reduced model

Scientific Reports, Год журнала: 2024, Номер 14(1)

Опубликована: Ноя. 2, 2024

Despite music's omnipresence, the specific neural mechanisms responsible for perceiving and anticipating temporal patterns in music are unknown. To study potential keeping time rhythmic contexts, we train a biologically constrained RNN, with excitatory (E) inhibitory (I) units, on seven different stimulus tempos (2–8 Hz) synchronization continuation task, standard experimental paradigm. Our trained RNN generates network oscillator that uses an input current (context parameter) to control oscillation frequency replicates key features of dynamics observed recordings monkeys performing same task. We develop reduced three-variable rate model analyze its dynamic properties. By treating our understanding mathematical structure oscillations as predictive, confirm dynamical found also RNN. neurally plausible reveals E-I circuit two distinct sub-populations, which one is tightly synchronized units.

Язык: Английский

---

Dynamic tracking of objects in the macaque dorsomedial frontal cortex

Nature Communications, Год журнала: 2025, Номер 16(1)

Опубликована: Янв. 2, 2025

Abstract A central tenet of cognitive neuroscience is that humans build an internal model the external world and use mental simulation to perform physical inferences. Decades human experiments have shown behaviors in many reasoning tasks are consistent with predictions from theory. However, evidence for defining feature – neural population dynamics reflect simulations states environment limited. We test hypothesis by combining a naturalistic ball-interception task, large-scale electrophysiology non-human primates, recurrent network modeling. find neurons monkeys’ dorsomedial frontal cortex (DMFC) represent task-relevant information about ball position multiplexed fashion. At level, activity pattern DMFC comprises low-dimensional embedding tracks both when it visible invisible, serving as substrate simulation. systematic comparison different classes task-optimized RNN models data provides further supporting hypothesis. Our findings provide environment.

Язык: Английский

---