Efficient recognition of emotions has attracted extensive research interest, which makes new applications in many fields possible, such as human-computer interaction, disease diagnosis, service robots, and so forth. Although existing work on sentiment analysis relying on sensors or unimodal methods performs well for simple contexts like business recommendation and facial expression recognition, it does far below expectations for complex scenes, such as sarcasm, disdain, and metaphors. In this article, we propose a novel two-stage multimodal learning framework, called AMSA, to adaptively learn correlation and complementarity between modalities for dynamic fusion, achieving more stable and precise sentiment analysis results. Specifically, a multiscale attention model with a slice positioning scheme is proposed to get stable quintuplets of sentiment in images, texts, and speeches in the first stage. Then a Transformer-based self-adaptive network is proposed to assign weights flexibly for multimodal fusion in the second stage and update the parameters of the loss function through compensation iteration. To quickly locate key areas for efficient affective computing, a patch-based selection scheme is proposed to iteratively remove redundant information through a novel loss function before fusion. Extensive experiments have been conducted on both machine weakly labeled and manually annotated datasets of self-made Video-SA, CMU-MOSEI, and CMU-MOSI. The results demonstrate the superiority of our approach through comparison with baselines.

Gestures, as a basic human feature, occupy an important position in human-computer interaction and other fields as well. In order to accurately recognize gestures and eliminate environmental interference, this paper proposes a gesture recognition matching method based on dynamic bones. This method uses the Mask R-CNN model and exponential filtering to identify and calibrate the key points of the hand. Through the segmentation and feature extraction of real-time frame images, the combined network is used to obtain stable and accurate hand bone key points. Based on the idea of constructing a Spatio-temporal graph convolutional network model based on ST-GCN, a skeletal point information gesture data set is constructed and sent to the network for training. Finally, template matching is used to realize gesture recognition. The experimental results show that the method can eliminate the environmental interference to the greatest extent, as well as the incomplete traditional data set, and the model accuracy defects caused by the lack of special samples. The recognition accuracy of the Chinese sign language database can reach 87.02%. Compared with previous researches on gesture recognition, it has improved accuracy and robustness.

Algorithms such as RRT (Rapidly exploring random tree), A* and their variants have been widely used in the field of robot path planning. A lot of work has shown that these detectors are unable to carry out effective and stable results for moving objects in high-dimensional space, which generate a large number of multi-dimensional corner points. Although some filtering mechanisms (such as splines and valuation functions) reduce the calculation scale, the chance of collision is increased, which is fatal to robots. In order to generate fewer but more effective and stable feature points, we propose a novel multi-scale positioning method to plan the motion of the high-dimensional target. First, a multi-scale feature extraction and refinement scheme for waypoint navigation and positioning is proposed to find the corner points that are more important to the planning, and gradually eliminate the unnecessary redundant points. Then, in order to obtain a stable planning effect, we balance the gradient of corner point classification detection to avoid over-optimizing some of them during the training phase. In addition, considering the maintenance cost of the robot in actual operation, we pay attention to the mechanism of anti-collision in the model design. Our approach can achieve a complete obstacle avoidance rate for high-dimensional space simulation and physical manipulators, and also work well in low-dimensional space for path planning. The experimental results demonstrate the superiority of our approach through a comparison with state-of-the-art models.

The World Wide Web needs reliable predictive capabilities to respond to changes in user behavior and usage patterns. Time series forecasting (TSF) is a key means to achieve this goal. In recent years, the large language models (LLMs) for TSF (LLM4TSF) have achieved good performance. However, there is a significant difference between pretraining corpora and time series data, making it hard to guarantee forecasting quality when directly applying LLMs to TSF; fine-tuning LLMs can mitigate this issue, but often incurs substantial computational overhead. Thus, LLM4TSF faces a dual challenge of prediction performance and compute overhead. To address this, we aim to explore a method for improving the forecasting performance of LLM4TSF while freezing all LLM parameters to reduce computational overhead. Inspired by in-context learning (ICL), we propose LVICL. LVICL uses our vector-injected ICL to inject example information into a frozen LLM, eliciting its in-context learning ability and thereby enhancing its performance on the example-related task (i.e., TSF). Specifically, we first use the LLM together with a learnable context vector adapter to extract a context vector from multiple examples adaptively. This vector contains compressed, example-related information. Subsequently, during the forward pass, we inject this vector into every layer of the LLM to improve forecasting performance. Compared with conventional ICL that adds examples into the prompt, our vector-injected ICL does not increase prompt length; moreover, adaptively deriving a context vector from examples suppresses components harmful to forecasting, thereby improving model performance. Extensive experiments demonstrate the effectiveness of our approach.

Meta-learning, or "learning to learn", enables machines to acquire general priors with minimal supervision and rapidly adapt to new tasks. Unlike traditional AI methods that approach each task from scratch using a fixed learning algorithm, meta-learning refines the learning algorithm itself through experience across various tasks, enhancing transferability and generalization. This is especially valuable when data collection is difficult or costly, allowing for effective learning from task sequences while reducing the dependency on extensive target domain data. Consequently, meta-learning has emerged as a promising field in machine learning. Although existing surveys provide valuable insights into meta-learning, they often present methods and applications in isolation and lack coverage of the latest advancements. Given the rapid growth of the field, a comprehensive survey is both necessary and challenging. Moreover, meta-learning algorithms often remain disconnected, with no unified framework to explain how they facilitate "learning to learn". This survey seeks to bridge that gap by systematizing meta-learning research, offering a thorough overview of strategies to enhance understanding. Additionally, the paper reviews over thirty representative meta-learning methods across models, tasks, and applications, analyzing their characteristics and challenges. To illustrate method performance, we evaluate more than fifteen models on six problems spanning thirteen scenarios, emphasizing the importance of selecting appropriate meta-learning approaches for practical applications.