Gang Wang

Flying insects demonstrate exceptional proficiency in detecting and pursuing conspecifics and prey within a cluttered environment, inspiring the development of computational models for small object detection. While existing bioinspired models are dedicated to resolving small moving instead of stationary object detection, few studies have systematically explored the role of visual motion in detection. Here, we developed a fly-inspired model on the basis of the hypothesis that combining visual motion features and luminance features is critical for small moving object detection. We thoroughly investigated the effect of feature combination under diverse stimulus conditions. Simulations indicated that the model exhibited hyperacute object detection, a capability not generally believed to emerge on the basis of motion detection. When tested with a moving background in realistic scenarios, the model demonstrated enhanced efficiency and robustness relative to models relying solely on luminance features. This enhancement was independent of whether visual motion was extracted by two- or three-arm motion detectors. The results suggested that small object detectors within the visual systems of flying insects could be optimally tuned to utilize the limited features inherent to tiny objects.

Understanding the brain requires modeling large-scale neural dynamics, where coarse-grained modeling of macroscopic brain behaviors is a powerful paradigm for linking brain structure to function with empirical data. However, the model inversion process remains computationally intensive and time-consuming, limiting research efficiency and medical deployment. In this work, we present a pipeline bridging coarse-grained brain modeling and advanced computing architectures. We introduce a dynamics-aware quantization framework that enables accurate low-precision simulation with maintained dynamical characteristics, thereby addressing the precision challenges inherent in the brain-inspired computing architecture. Furthermore, to exploit hardware capabilities, we develop hierarchical parallelism mapping strategies tailored for brain-inspired computing chips and GPUs. Experimental results demonstrate that the deployed low-precision models maintain high functional fidelity while achieving tens to hundreds-fold acceleration over commonly used CPUs. This work provides essential computational infrastructures for modeling macroscopic brain dynamics and extends the application of brain-inspired computing to scientific computing in neuroscience and medicine.

Spiking Neural Networks (SNNs), inspired by neuroscience principles, have gained attention for their energy efficiency. However, directly trained SNNs lag behind Artificial Neural Networks (ANNs) in accuracy for complex tasks like object detection due to the limited information capacity of binary spike feature maps. To address this, we propose BD-SNN, a new directly trained SNN equipped with Bidirectional Dynamic Threshold neurons (BD-LIF). BD-LIF neurons emit +1 and –1 spikes and dynamically adjust their thresholds, enhancing the network's information encoding capacity and activation efficiency. Our BD-SNN incorporates two new all-spike residual blocks, BD-Block1 and BD-Block2, for efficient information extraction and multi-scale feature fusion, respectively. Experiments on the COCO and Gen1 datasets demonstrate that BD-SNN improves accuracy by 3.1% and 2.8% compared to the state-of-the-art EMS-YOLO method, respectively, validating BD-SNN's superior performance across diverse input scenarios. Project will be available at https://github.com/Ganpei576/BD-SNN.

Traditional object detection methods usually underperform when locating tiny or small drones against complex backgrounds, since the appearance features of the targets and the backgrounds are highly similar. To address this, inspired by the magnocellular motion processing mechanisms, we proposed to utilize the spatial–temporal characteristics of the flying drones based on spiking neural networks, thereby developing the Magno-Spiking Neural Network (MG-SNN) for drone detection. The MG-SNN can learn to identify potential regions of moving targets through motion saliency estimation and subsequently integrates the information into the popular object detection algorithms to design the retinal-inspired spiking neural network module for drone motion extraction and object detection architecture, which integrates motion and spatial features before object detection to enhance detection accuracy. To design and train the MG-SNN, we propose a new backpropagation method called Dynamic Threshold Multi-frame Spike Time Sequence (DT-MSTS), and establish a dataset for the training and validation of MG-SNN, effectively extracting and updating visual motion features. Experimental results in terms of drone detection performance indicate that the incorporation of MG-SNN significantly improves the accuracy of low-altitude drone detection tasks compared to popular small object detection algorithms, acting as a cheap plug-and-play module in detecting small flying targets against complex backgrounds.

Sensitivity to motion direction is a feature of visual neurons that is essential for motion perception. Recent studies have suggested that direction selectivity is re-established at multiple stages throughout the visual hierarchy, which contradicts the traditional assumption that direction selectivity in later stages largely derives from that in earlier stages. By recording laminar responses in areas 17 and 18 of anesthetized cats of both sexes, we aimed to understand how direction selectivity is processed and relayed across 2 successive stages: the input layers and the output layers within the early visual cortices. We found a strong relationship between the strength of direction selectivity in the output layers and the input layers, as well as the preservation of preferred directions across the input and output layers. Moreover, direction selectivity was enhanced in the output layers compared to the input layers, with the response strength maintained in the preferred direction but reduced in other directions and under blank stimuli. We identified a direction-tuned gain mechanism for interlaminar signal transmission, which likely originated from both feedforward connections across the input and output layers and recurrent connections within the output layers. This direction-tuned gain, coupled with nonlinearity, contributed to the enhanced direction selectivity in the output layers. Our findings suggest that direction selectivity in later cortical stages partially inherits characteristics from earlier cortical stages and is further refined by intracortical connections.