Point cloud data has been widely used in many fields, and processing them using deep learning methods has become a popular research topic. However, the irregularity and unordered data structure of point cloud makes it necessary to design neural networks that are different from those used in image and natural language processing to accommodate its characteristics. In this paper, we propose a novel network called the Point Distance Mask Attention Network. This network utilizes the position relationship between the centroid and the generated neighborhood to weight the features of each point, making the network more focused on the features close to the centroid. Additionally, we use a masking operation to ensure the permutation invariance of the network when using the ball query method to find the neighborhood. Furthermore, we propose a residual-like connection in the network architecture, which achieves better results without changing the network feature extractor and depth. We evaluated the network model on the ModelNet40 dataset and achieved an accuracy of 93.7%. Our experiments show that our network achieves better results than the original baseline network by 1.1%, with almost the same parameters and inference time.
Spiking neural networks have the nature of high efficiency, energy saving, and bio-interpretability. They communicate through sparse and asynchronous spikes, so they have received extensive attention in the field of neuromorphic engineering and brain-like computing. At present, the commonly used encoding methods are mainly single-rate encoding and temporal encoding. However, rate encoding cannot make use of the time information in the spike train, which has high energy consumption. Temporal encoding limits the computing power of neurons and will produce dead neurons. Moreover, it is critical to find effective solutions that reduce network complexity and improve energy efficiency while maintaining high accuracy. Therefore, we propose a hybrid coding method based on rate coding and temporal coding to solve the limitation of single coding. We propose an adaptive online pruning strategy based on hybrid coding. In this pruning strategy, 100 neurons are pruned out of the 200-neuron network, which reduces the network size and obtains a more compact network structure. The memory capacity is reduced by 1.9×, the energy efficiency is increased by 2.4×, and the classification accuracy is reduced by less than 0.5%.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.