As an important neural network model, the continuous attractor neural network (CANN) demonstrates unique advantages in simulating and explaining the representation and storage mechanisms of continuous variables (such as position, direction, etc.) in the brain. However, in the real brain network, the connections between neurons are not only random or equal, but regulated by various factors. The coding methods and interactions of multiple neuron groups form the collective behavior of the complex brain network. To study the role of complex networks in the brain, this research proposes a selective coupling network model based on CANN, which operates under the dynamic balance between excitatory and inhibitory inputs. In this study, we have investigated the interaction between two classes of CANNs with distinct selective preferences for motion direction recognition under fast excitation and inhibition (E-I) balance. Based on simulation results and theoretical analysis, we found that the fast E-I balance located in the brain can provide indirect linking effects among multiple different networks. The indirect mutual inhibition of the coupled network through E-I balance can promote the accuracy and stability of the network response to a given position. This suggests that a dynamic system of collaborative work activity is possible between different neural networks. We hope that this work could provide insightful references for the development of multi-network coupling in brain-like computing.

Medical professionals rely on MRI scans to evaluate brain tumors, a task prone to fatigue and subjective interpretation. This study enhances the YOLOv10 model by integrating two innovative modules: AvConv module and Bidirectional Pathway, both designed to improve detection accuracy. Experiments were conducted using two Kaggle datasets (Dataset 1: 1116 images; Dataset 2: 153 images), both annotated for tumor presence. The proposed model was compared against eight mainstream object detection methods. Results demonstrate that the proposed AvConv module achieves mAP50 of 88.5% and mAP50-95 of 77.7%, which represent relative improvements of 0.3% and 3.1%, respectively, over the baseline YOLOv10 achieves mAP50 of 88.2% and mAP50-95 of 74.6%. Meanwhile, the Bi-directional detection model reaches 90.3% mAP50 and 83.5% mAP50-95, corresponding to 2.1% and 8.9% gains relative to YOLOv10. These enhancements validate the effectiveness of the introduced modules in advancing detection accuracy while maintaining competitive inference efficiency, supporting the role of customized deep learning architectures in reliable, automated brain tumor diagnosis.