Relaxing in Warped Spaces: Generalized Hierarchical and Modular Dynamical Neural Network

arXiv (Cornell University) · 2026-04-12

We propose a dynamical neural network model with a hierarchical and modular structure. The network architecture can be derived by minimizing an energy function that is originally designed based on two kinds of neurons with quite different time constants. It has multiple subspaces that are spanned by neural parameters employed in the energy function, and adjacent subspaces are related to each other with a layered internetwork. Each internetwork further consists of a pair of a forward subnet and a backward one, and signals flowing through these subnets determine total dynamics of the network. The model can operate in either a learning or an association mode. In the learning mode, when periodic signals equivalent to repetitive neuronal bursting are suitably applied to input ports in all subspaces, mapping relationships corresponding to those input signals are eventually formed in internetworks between subspaces. Various two-dimensional mapping relationships between subspaces can be shaped by employing an appropriate set of periodic input signals with different frequencies based on the same mechanism as a Lissajous curve. The model in the association mode provides an overall framework such that state variables inside the network individually relax in warped spaces, each of which has been designed as favorable for a (or some) state variable(s). The association mode is further classified into two modes; unconstrained and constrained. In the latter mode, for instance, when a sufficiently slow periodic trajectory is set as an input, a warped output trajectory appears in each subspace as if imaginary layered networks with the inverse mapping relationships to existing forward subnets' were located hierarchically from outside to inside. These results suggest that a certainty/uncertainty relation exists between an input trajectory and an output trajectory.

Relaxing in Warped Spaces: Generalized Hierarchical and Modular Dynamical Neural Network

arXiv (Cornell University) · 2026-04-12

We propose a dynamical neural network model with a hierarchical and modular structure. The network architecture can be derived by minimizing an energy function that is originally designed based on two kinds of neurons with quite different time constants. It has multiple subspaces that are spanned by neural parameters employed in the energy function, and adjacent subspaces are related to each other with a layered internetwork. Each internetwork further consists of a pair of a forward subnet and a backward one, and signals flowing through these subnets determine total dynamics of the network. The model can operate in either a learning or an association mode. In the learning mode, when periodic signals equivalent to repetitive neuronal bursting are suitably applied to input ports in all subspaces, mapping relationships corresponding to those input signals are eventually formed in internetworks between subspaces. Various two-dimensional mapping relationships between subspaces can be shaped by employing an appropriate set of periodic input signals with different frequencies based on the same mechanism as a Lissajous curve. The model in the association mode provides an overall framework such that state variables inside the network individually relax in warped spaces, each of which has been designed as favorable for a (or some) state variable(s). The association mode is further classified into two modes; unconstrained and constrained. In the latter mode, for instance, when a sufficiently slow periodic trajectory is set as an input, a warped output trajectory appears in each subspace as if imaginary layered networks with the inverse mapping relationships to existing forward subnets' were located hierarchically from outside to inside. These results suggest that a certainty/uncertainty relation exists between an input trajectory and an output trajectory.

The role of attention in multi attribute decision making

Nature Communications · 2025-11-29

Real-life decisions typically involve multiple options, each with multiple attributes affecting value. In such complex situations, sequential shifts of attention to specific options and attributes are thought to guide the decision process. Using a task that allowed us to monitor attention during such multi-attribute decisions, we recorded decision-related signals in pre-supplementary motor area neurons from two male macaques. Attention influences activity in these neurons through two mechanisms. First, attention enhances the activity of neurons representing the currently sampled option, independent of its value, without fully suppressing the representation of other options. Second, attention up-regulates the gain of information integration towards the evolving value estimate for the attended option. In contrast, we found no evidence for a third suggested mechanism, in which only the attended option is represented. Instead, attention influences the ongoing parallel information accumulation and competition process by modulating the strength of the value information that drives this circuit. How attention facilitates complex decisions requiring the consideration of multiple options with multiple attributes remains unclear. Here, the authors found that the brain represents multiple options simultaneously and that attention modulates the representation of their value.

Overt Visual Attention in the Formation of Preference Between Complex Lottery Options

Computational Brain & Behavior · 2025-05-07 · 1 citations

Event-driven figure-ground organisation model for the humanoid robot iCub

Nature Communications · 2025-02-21 · 5 citations

Figure-ground organisation is a perceptual grouping mechanism for detecting objects and boundaries, essential for an agent interacting with the environment. Current figure-ground segmentation methods rely on classical computer vision or deep learning, requiring extensive computational resources, especially during training. Inspired by the primate visual system, we developed a bio-inspired perception system for the neuromorphic robot iCub. The model uses a hierarchical, biologically plausible architecture and event-driven vision to distinguish foreground objects from the background. Unlike classical approaches, event-driven cameras reduce data redundancy and computation. The system has been qualitatively and quantitatively assessed in simulations and with event-driven cameras on iCub in various scenarios. It successfully segments items in diverse real-world settings, showing comparable results to its frame-based version on simple stimuli and the Berkeley Segmentation dataset. This model enhances hybrid systems, complementing conventional deep learning models by processing only relevant data in Regions of Interest (ROI), enabling low-latency autonomous robotic applications. Inspired by the primate visual system, this work implements an event-driven, bio-inspired architecture for figure-ground segmentation on the neuromorphic robot iCub, bridging neuromorphic algorithms and software. Its performance is benchmarked on the Berkeley Segmentation Data Set and validated in real-world scenarios.

Neural correlates from the middle temporal gyrus via machine learning decode decision making variables during gambling in humans

2025-03-19

The ability to make good decisions is contingent upon extracting the most relevant information from one’s contextual surroundings. The middle temporal gyrus lies at the nexus of visual processing streams and is hypothesized to play a crucial role in the semantic processing of visual information. In this study, stereoelectroencephalography measurements from the middle temporal gyrus were used to train random forest classifiers to decode betting behavior in a financial-decision making task for 10 subjects. Spectral features from the middle temporal gyrus were computed. The performance of these subject-specific classifiers ranged between 75% and 98% accuracy when predicting human betting behavior and provide evidence that the middle temporal gyrus encodes information related to decision-making. This study is the first to use stereoelectroencephalography to explore the functionality of the middle temporal gyrus in decision-making.

Event-Driven Figure-Ground Organisation model for the humanoid robot iCub

Research Square · 2024-02-05

An Exploratory Study of Large-Scale Brain Networks during Gambling Using SEEG

Brain Sciences · 2024-07-31 · 2 citations

Decision-making is a cognitive process involving working memory, executive function, and attention. However, the connectivity of large-scale brain networks during decision-making is not well understood. This is because gaining access to large-scale brain networks in humans is still a novel process. Here, we used SEEG (stereoelectroencephalography) to record neural activity from the default mode network (DMN), dorsal attention network (DAN), and frontoparietal network (FN) in ten humans while they performed a gambling task in the form of the card game, "War". By observing these networks during a decision-making period, we related the activity of and connectivity between these networks. In particular, we found that gamma band activity was directly related to a participant's ability to bet logically, deciding what betting amount would result in the highest monetary gain or lowest monetary loss throughout a session of the game. We also found connectivity between the DAN and the relation to a participant's performance. Specifically, participants with higher connectivity between and within these networks had higher earnings. Our preliminary findings suggest that connectivity and activity between these networks are essential during decision-making.

An Event-Based Implementation of Saliency-Based Visual Attention for Rapid Scene Analysis

IEEE Transactions on Cognitive and Developmental Systems · 2024-10-22 · 2 citations

Selective attention is an essential mechanism to filter sensory input and to select only its most important components, allowing the capacity-limited cognitive structures of the brain to process them in detail. The saliency map model, originally developed to understand the process of selective attention in the primate visual system, has also been extensively used in computer vision. Due to the wide-spread use of frame-based video, this is how dynamic input from non-stationary scenes is commonly implemented in saliency maps. However, the temporal structure of this input modality is very different from that of the primate visual system. Retinal input to the brain is massively parallel, local rather than frame-based, asynchronous rather than synchronous, and transmitted in the form of discrete events, neuronal action potentials (spikes). These features are captured by event-based cameras. We show that a computational saliency model can be obtained organically from such vision sensors, at minimal computational cost. We assess the performance of the model by comparing its predictions with the distribution of overt attention (fixations) of human observers, and we make available an event-based dataset that can be used as ground truth for future studies.

An event-based implementation of saliency-based visual attention for rapid scene analysis

arXiv (Cornell University) · 2024-01-10

Selective attention is an essential mechanism to filter sensory input and to select only its most important components, allowing the capacity-limited cognitive structures of the brain to process them in detail. The saliency map model, originally developed to understand the process of selective attention in the primate visual system, has also been extensively used in computer vision. Due to the wide-spread use of frame-based video, this is how dynamic input from non-stationary scenes is commonly implemented in saliency maps. However, the temporal structure of this input modality is very different from that of the primate visual system. Retinal input to the brain is massively parallel, local rather than frame-based, asynchronous rather than synchronous, and transmitted in the form of discrete events, neuronal action potentials (spikes). These features are captured by event-based cameras. We show that a computational saliency model can be obtained organically from such vision sensors, at minimal computational cost. We assess the performance of the model by comparing its predictions with the distribution of overt attention (fixations) of human observers, and we make available an event-based dataset that can be used as ground truth for future studies.