Reconstruction-based anomaly detection methods often struggle with invariant reconstruction of abnormal regions and the unintended reconstruction of novel anomalies. To address these limitations, this study proposes a novel guided training and reconstruction framework (SGAD) to enhance anomaly reconstruction quality. The proposed approach integrates a training paradigm built on target images and fusion loss, along with target-guided and secondary reconstruction strategies utilizing a diffusion model, achieving superior anomaly detection performance. Additionally, a new DTY anomaly detection dataset is introduced to benchmark the approach. Extensive experiments were conducted on the DTY and MVTec datasets, demonstrating that SGAD achieves state-of-the-art performance, with mean scores of 93.7% I-AUROC and 86.2% P-AUROC. These results highlight the effectiveness and robustness of SGAD in addressing complex anomaly detection challenges, underscoring its potential for deployment in practical production environments.

The segmentation of dense point clouds from industrial LiDAR scans presents challenges in computational overhead and VRAM usage, hindering the development of automated fast measurement systems. To address this, we propose EPNet, an efficient model for part segmentation of dense point clouds. EPNet employs a U-Net-like architecture with skip connections to merge original and recovered features, enhancing local feature extraction via KNN and cosine similarity. Factorization-dimensionality-reduction module based on self-attention overcomes the limitations of trilinear interpolation in feature recovery, improving both local and global feature fusion. In experiments on the LVPC dataset of dense vehicle point clouds, EPNet outperforms models from the past three years, achieving a 1.7% accuracy improvement and a 9.7% increase in average Instance IoU compared to PointNet++. EPNet also achieves a single-file inference time of under 1 second while requiring minimal GPU VRAM resources, demonstrating its potential for real-world industrial high-precision fast automated measurements. The code is available at https://github.com/duskNNNN/EPNet.

Action segmentation aims to assign accurate action labels to each frame in a video, and plays an important role in behavior understanding, video surveillance, and human-computer interaction. However, existing methods mainly focus on modeling temporal dependencies while paying little attention to spatial information within frames. This limitation leads to insufficient representation of key regions that are crucial for recognizing complex hand movements or human-object interactions. In addition, the transitions between action and background frames are often ambiguous, which further reduces segmentation accuracy. To address these challenges, we propose SAAB, a new action segmentation framework that combines Spatial Attention Pooling and an Action-Background Classifier. The spatial attention pooling module performs weighted fusion of features from different spatial regions to enhance intra-frame structural awareness. At the same time, the action-background classifier explicitly separates action frames from background frames, improving the reliability of frame-level predictions. Extensive experiments on multiple benchmark datasets show that SAAB clearly improves segmentation performance and can be easily integrated into various existing segmentation models, demonstrating its effectiveness and generality.

Reinforcement learning has been widely used in explaining animal behavior. In reinforcement learning, the agent learns the value of the states in the task, collectively constituting the task state space, and uses the knowledge to choose actions and acquire desired outcomes. It has been proposed that the orbitofrontal cortex (OFC) encodes the task state space during reinforcement learning. However, it is not well understood how the OFC acquires and stores task state information. Here, we propose a neural network model based on reservoir computing. Reservoir networks exhibit heterogeneous and dynamic activity patterns that are suitable to encode task states. The information can be extracted by a linear readout trained with reinforcement learning. We demonstrate how the network acquires and stores task structures. The network exhibits reinforcement learning behavior and its aspects resemble experimental findings of the OFC. Our study provides a theoretical explanation of how the OFC may contribute to reinforcement learning and a new approach to understanding the neural mechanism underlying reinforcement learning.

Individuals in large groups respond to the movements and positions of their neighbors by following a set of interaction rules. These rules are central to understanding the mechanisms of collective motion. However, whether individuals actually use these rules to guide their movements remains untested. Here we show that the real-time movements of individual fish can be directly predicted from their neighbors’ motion. We train a recurrent neural network to predict the trajectories of individual fish from input signals. The inputs are projected to the recurrent network as time series representing the movements and positions of neighboring fish. By comparing the data output from the model with the target fish’s trajectory, we provide direct evidence that individuals guide their movements via interaction rules. Because the error between the model output and actual trajectory changes when the fish perceive a noxious contaminant, the model is potentially applicable to water quality monitoring.