Edit/Keich
Damage to the part of the brain that processes visual information — the subtemporal (IT) cortex — can be devastating, especially for adults. Those affected may lose the ability to read (a condition called dyslexia) or the ability to recognize faces (facial agnosia) or objects (agnosia), which doctors are currently unable to do.
More accurate models of the visual system could help neuroscientists and clinicians develop better treatments for these diseases.
Recently, researchers at Carnegie Mellon University (CMU) developed a computational model, the Interactive Terrain Network (ITN), that allows them to simulate it's spatial organization or topography and learn more about how adjacent clusters of brain tissue organize and interact. This can also help them understand how damage to the area affects the ability to recognize faces, objects, and scenes.
The study, titled "A connectivity-constrained computational account of topographic organization in primate high-level visual cortex," was published in PNAS on January 18, 2022.
Dr. Nicholas M. Blauch said the paper may help cognitive neuroscientists answer long-standing questions about how different parts of the brain work together.
"We've long wondered if we should think of the network of regions in the brain that respond to faces as a separate entity that is only used to recognize faces, or should we use it as part of the neural structure of target recognition," Blauch said. "We're trying to solve this problem using a computational model that assumes this simpler, more general organization, and see if this model can explain the specializations we see in the brain by learning to perform tasks."
To do this, the researchers developed a deep learning model: interactive terrain networks (ITNs), which have additional features of biological brain connections, assuming that the model can reveal the spatial organization or terrain of IT.
Interactive terrain network
ITN, a framework for computational modeling of the visual cortex, in particular its functional terrain organization. An ITN model is defined as a neural network model that: (1) optimizes to perform natural tasks; (2) constrains connections in a biologically reasonable way to produce functional organization.
In this work, a form of ITN is introduced, which is divided into three parts: an encoder that approximates the early visual cortex, an interactive terrain (IT) layer that approximates the subtemporal cortex, and a readout mechanism for one or more downstream tasks. The goal of the encoder is to extract general visual features that describe the visual world along the dimensions that support a wide range of downstream readout tasks.
The researchers' primary modeling focus was on the IT layer, which consists of a series of biologically constrained pairs of cyclic layers. For simplicity of calculation, these constraints are not modeled in the encoder.
Simulation results of a particular ITN model are first shown, referred to as the main model or "E/I-EFF-RNN", suggesting that it has independent neurons responsible for excitation and inhibition (E/I). Feed-forward connections are strictly excitatory (EFF) limitations, and time loop processing is mediated by learning lateral connections (RNNs). In addition, the model uses the ResNet-50 encoder, which is pre-trained on large datasets, including multiple categories from objects, faces, and scene domains, and is used as a feature extractor after pre-training It provides input to three-region IT with separate pIT, cIT, and aIT regions.
Illustration: An interactive terrain network produces a hierarchical domain-level organization. (Source: Thesis)
After training, the model performed well in various fields, achieving 86.4% classification accuracy in the face domain, 81.8% in the object domain, and 65.9% in the scene domain. Cross-domain performance differences are unlikely to be the product of a particular architecture, as they can be seen in a variety of DCNNs, which reflects the inherent difficulty of each task due to variability within and between categories within and between categories in each domain of a given image set.
To further confirm the functional significance of topographic organization, the researchers analyzed the spatial organization of read weights from aIT to the local category readout layer. The study found a large positive correlation between the average reading weight and the average response per domain (all rs > 0.7, all Ps
Damage analysis
Next, the researchers performed a series of damage analyses in the model to compare with neuropsychological data for facial and object recognition. First, focal lesions were done.
Illustration: Damage results in an ITN model. (Source: Thesis)
Studies have shown that focal lesions centered on each domain lead to unusually severe recognition of that domain, while there are also minor but significant deficiencies in other domains. For such lesions, defects in all domains are significant (all Ps
Partial but not all, damage to non-preferred areas due to focal lesions may be due to imperfect or non-circular topographically functional tissue. Importantly, these more dispersed effects of lesions suggest that functional tissue, while highly specialized, is not strictly modular; damage to those units that claim to be part of a given module (e.g., for facial recognition) still affects object recognition (albeit to a weaker degree).
"There is some residual damage to other areas," Blauch said. "It's small compared to preferred domains, but it shows us that the specialization in these networks can be strong, but it's also a bit mixed." Combined with the general principles adopted by the entire system, this means that it may be considered a system with internal specialization, rather than a collection of independent modules."
A universal, flexible system may be more capable of recombining after injury, as we have seen in children, where visual function is essentially restored after impaired in infancy compared to adults with similar impairments.
Limitations and future directions
The current work only deals with the topographic organization of the high-level representation. Modeling topology organization in convolutional layers is a particular challenge for ITN frameworks. These architectures and other biologically plausible variants are an exciting opportunity to examine topographic organization from connection-based constraints.
Related to this, while the ITN has an advantage in interpreting hierarchical topographic organization due to the limitations of interregional space, it has not yet satisfactorily specifically explained some aspects of hierarchical representation transformations, increasing the invariance of 3D rotations. Future work urgently requires extending the ITN framework to more powerful computing architectures, training environments, and learning rules, rather than delegating this computing power to different encoders.
There are some differences between the ITN model and the overall representation space of primate IT. Comparing different ITN models in more detail to explain the effects of the IT cortex quantitatively and qualitatively is an exciting route for future research.
While the work improves biological plausibility with previous work, by combining routing constraints, separation of excitation and inhibition, and excitatory connections between regions, additional biological details may be important for the calculation and organization of the visual cortex. Future work may consider incorporating details.
The research work has important implications for cognitive neuroscience, providing a general description of the development of the field of topographic function specialization, and for computational neuroscience, by showing how well-known biological details can be incorporated into neural network models to explain empirical findings.
Reference: https://medicalxpress.com/news/2022-02-neuroscientists-deep-simulate-brain-topography.html
Artificial Intelligence × [ Biological Neuroscience Mathematics Physics Materials ]
"ScienceAI" focuses on the intersection and integration of artificial intelligence with other cutting-edge technologies and basic sciences.
Welcome to follow the stars and click Likes and Likes and Are Watching in the bottom right corner.