An official website of the United States government
Recurrent neural networks (RNNs) trained on a diverse ensemble of cognitive tasks, as described by Yang et al. (2019); Khona et al. (2023), have been shown to exhibit functional modularity, where neurons organize into discrete functional clusters, each specialized for specific shared computational subtasks. However, these RNNs do not demonstrate anatomical modularity, where these functionally specialized clusters also have a distinct spatial organization. This contrasts with the human brain which has both functional and anatomical modularity. Is there a way to train RNNs to make them more like brains in this regard? We apply a recent machine learning method, brain-inspired modular training (BIMT), to encourage neural connectivity to be local in space. Consequently, hidden neuron organization of the RNN forms spatial structures reminiscent of those of the brain: spatial clusters which correspond to functional clusters. Compared to standard L1 regularization and absence of regularization, BIMT exhibits superior performance by optimally balancing between task performance and sparsity. This balance is quantified both in terms of the number of active neurons and the cumulative wiring length. In addition to achieving brain-like organization in RNNs, our findings also suggest that BIMT holds promise for applications in neuromorphic computing and enhancing the interpretability of neural network architectures.
more »
« less
Neural Sculpting: Uncovering hierarchically modular task structure in neural networks through pruning and network analysis
Natural target functions and tasks typically exhibit hierarchical modularity – they can be broken down into simpler sub-functions that are organized in a hierarchy. Such sub-functions have two important features: they have a distinct set of inputs (input-separability) and they are reused as inputs higher in the hierarchy (reusability). Previous studies have established that hierarchically modular neural networks, which are inherently sparse, offer benefits such as learning efficiency, generalization, multi-task learning, and transfer. However, identifying the underlying sub-functions and their hierarchical structure for a given task can be challenging. The high-level question in this work is: if we learn a task using a sufficiently deep neural network, how can we uncover the underlying hierarchy of sub-functions in that task? As a starting point, we examine the domain of Boolean functions, where it is easier to determine whether a task is hierarchically modular. We propose an approach based on iterative unit and edge pruning (during training), combined with network analysis for module detection and hierarchy inference. Finally, we demonstrate that this method can uncover the hierarchical modularity of a wide range of Boolean functions and two vision tasks based on the MNIST digits dataset.
more »
« less
- Award ID(s):
- 2039741
- PAR ID:
- 10505215
- Publisher / Repository:
- Advances in Neural Information Processing Systems
- Date Published:
- Journal Name:
- Advances in neural information processing systems
- ISSN:
- 1049-5258
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Neural dynamics can reflect intrinsic dynamics or dynamic inputs, such as sensory inputs or inputs from other brain regions. To avoid misinterpreting temporally structured inputs as intrinsic dynamics, dynamical models of neural activity should account for measured inputs. However, incorporating measured inputs remains elusive in joint dynamical modeling of neural-behavioral data, which is important for studying neural computations of behavior. We first show how training dynamical models of neural activity while considering behavior but not input or input but not behavior may lead to misinterpretations. We then develop an analytical learning method for linear dynamical models that simultaneously accounts for neural activity, behavior, and measured inputs. The method provides the capability to prioritize the learning of intrinsic behaviorally relevant neural dynamics and dissociate them from both other intrinsic dynamics and measured input dynamics. In data from a simulated brain with fixed intrinsic dynamics that performs different tasks, the method correctly finds the same intrinsic dynamics regardless of the task while other methods can be influenced by the task. In neural datasets from three subjects performing two different motor tasks with task instruction sensory inputs, the method reveals low-dimensional intrinsic neural dynamics that are missed by other methods and are more predictive of behavior and/or neural activity. The method also uniquely finds that the intrinsic behaviorally relevant neural dynamics are largely similar across the different subjects and tasks, whereas the overall neural dynamics are not. These input-driven dynamical models of neural-behavioral data can uncover intrinsic dynamics that may otherwise be missed.more » « less
-
Abstract The concept of control is crucial for effectively understanding and applying biological network models. Key structural features relate to control functions through gene regulation, signaling, or metabolic mechanisms, and computational models need to encode these. Applications often focus on model-based control, such as in biomedicine or metabolic engineering. In a recent paper, the authors developed a theoretical framework of modularity in Boolean networks, which led to a canonical semidirect product decomposition of these systems. In this paper, we present an approach to model-based control that exploits this modular structure, as well as the canalizing features of the regulatory mechanisms. We show how to identify control strategies from the individual modules, and we present a criterion based on canalizing features of the regulatory rules to identify modules that do not contribute to network control and can be excluded. For even moderately sized networks, finding global control inputs is computationally challenging. Our modular approach leads to an efficient approach to solving this problem. We apply it to a published Boolean network model of blood cancer large granular lymphocyte (T-LGL) leukemia to identify a minimal control set that achieves a desired control objective.more » « less
-
Abstract The brain’s modular functional organization facilitates adaptability. Modularity has been linked with a wide range of cognitive abilities such as intelligence, memory, and learning. However, much of this work has (1) considered modularity while a participant is at rest rather than during tasks conditions and/or (2) relied primarily on lab-based cognitive assessments. Thus, the extent to which modularity can provide information about real-word behavior remains largely unknown. Here, we investigated whether functional modularity during resting-state and task-based fMRI was associated with academic learning (measured by GPA) and ability (measured by PSAT) in a large sample of high school students. Additional questions concerned the extent to which modularity differs between rest and task conditions, and across spatial scales. Results indicated that whole-brain modularity during task conditions was significantly associated with academic learning. In contrast to prior work, no such associations were observed for resting-state modularity. We further showed that differences in modularity between task conditions and resting-state varied across spatial scales. Taken together, the present findings inform how functional brain network modularity – during task conditions and while at rest – relate to a range of cognitive abilities.more » « less
-
The structure of neural circuitry plays a crucial role in brain function. Previous studies of brain organization generally had to trade off between coarse descriptions at a large scale and fine descriptions on a small scale. Researchers have now reconstructed tens to hundreds of thousands of neurons at synaptic resolution, enabling investigations into the interplay between global, modular organization, and cell type-specific wiring. Analyzing data of this scale, however, presents unique challenges. To address this problem, we applied novel community detection methods to analyze the synapse-level reconstruction of an adult femaleDrosophila melanogasterbrain containing >20,000 neurons and 10 million synapses. Using a machine-learning algorithm, we find the most densely connected communities of neurons by maximizing a generalized modularity density measure. We resolve the community structure at a range of scales, from large (on the order of thousands of neurons) to small (on the order of tens of neurons). We find that the network is organized hierarchically, and larger-scale communities are composed of smaller-scale structures. Our methods identify well-known features of the fly brain, including its sensory pathways. Moreover, focusing on specific brain regions, we are able to identify subnetworks with distinct connectivity types. For example, manual efforts have identified layered structures in the fan-shaped body. Our methods not only automatically recover this layered structure, but also resolve finer connectivity patterns to downstream and upstream areas. We also find a novel modular organization of the superior neuropil, with distinct clusters of upstream and downstream brain regions dividing the neuropil into several pathways. These methods show that the fine-scale, local network reconstruction made possible by modern experimental methods are sufficiently detailed to identify the organization of the brain across scales, and enable novel predictions about the structure and function of its parts. Significance StatementThe Hemibrain is a partial connectome of an adult femaleDrosophila melanogasterbrain containing >20,000 neurons and 10 million synapses. Analyzing the structure of a network of this size requires novel and efficient computational tools. We applied a new community detection method to automatically uncover the modular structure in the Hemibrain dataset by maximizing a generalized modularity measure. This allowed us to resolve the community structure of the fly hemibrain at a range of spatial scales revealing a hierarchical organization of the network, where larger-scale modules are composed of smaller-scale structures. The method also allowed us to identify subnetworks with distinct cell and connectivity structures, such as the layered structures in the fan-shaped body, and the modular organization of the superior neuropil. Thus, network analysis methods can be adopted to the connectomes being reconstructed using modern experimental methods to reveal the organization of the brain across scales. This supports the view that such connectomes will allow us to uncover the organizational structure of the brain, which can ultimately lead to a better understanding of its function.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Proceeding:
- The DOI is not currently available.
