Search

In response to the issue that traditional lung nodule detection models cannot dynamically optimize and update with the increase of new data, a new lung nodule detection model—task incremental meta-learning model (TIMLM) is proposed. This model comprises of two loops: the inner loop imposes incremental learning regularization update constraints, while the outer loop employs a meta-update strategy to sample old and new knowledge and learn a set of generalized parameters that adapt to old and new data. Under the condition that the main structure of the model is not changed as much as possible, it preserves the old knowledge that was learned previously. Experimental verification on the publicly available lung dataset showed that, compared with traditional deep network models and mainstream incremental models, TIMLM has greatly improved in terms of accuracy, sensitivity, and other indicators, demonstrating good continuous learning and anti-forgetting capabilities.

Stock price prediction has always been a focal topic in financial research and quantitative investment. Currently, most deep learning models for stock price prediction are based on batch learning settings, which require prior knowledge of the training dataset. These models are not scalable for real-time data stream prediction, and their performance decreases when the data distribution dynamically changes. To address the issue of poor prediction accuracy for non-stationary stock price data in existing research, this paper proposes an online stock price prediction model (Increformer) based on incremental learning and continuous attention mechanism. By leveraging continuous self-attention mechanism to capture the temporal dependencies among feature variables and employing continuous normalization mechanism to handle non-stationary data, the model enhances prediction accuracy through the incremental training strategy based on elastic weighting consolidation to acquire new knowledge from the data stream. Five public datasets are selected from the stock index and individual stock price sequences in the stock market for experiments. Experimental results demonstrate that Increformer effectively extracts temporal information and feature dimension correlation from the data, thus improving the prediction performance of stock prices. Additionally, ablation experiments are conducted to evaluate the effects and necessity of the continuous normalization mechanism, continuous attention mechanism, and incremental training strategy, thereby verifying the accuracy and generalizability of the proposed model. Increformer can effectively capture the trends and fluctuations of stock price series.

Defect detection constitutes one of the most crucial processes in industrial production. With a continuous increase in the number of defect categories and samples, the defect detection model underpinned by deep learning finds it challenging to expand to new categories, and the accuracy and real-time performance of product defect detection are also confronted with severe challenges. This paper addresses the problem of insufficient detection accuracy of existing lightweight models on resource-constrained edge devices by presenting a new lightweight YoloV5 model, which integrates four modules, SCDown, GhostConv, RepNCSPELAN4, and ScalSeq. Here, this paper abbreviates it as SGRS-YoloV5n. Through the incorporation of these modules, the model notably enhances feature extraction and computational efficiency while reducing the model size and computational load, making it more conducive for deployment on edge devices. Furthermore, a cloud-edge collaborative defect detection system is constructed to improve detection accuracy and efficiency through initial detection by edge devices, followed by additional inspection by cloud servers. An incremental learning mechanism is also introduced, enabling the model to adapt promptly to new defect categories and update its parameters accordingly. Experimental results reveal that the SGRS-YoloV5n model exhibits superior detection accuracy and real-time performance, validating its value and stability for deployment in resource-constrained environments. This system presents a novel solution for achieving efficient and accurate real-time defect detection. [ABSTRACT FROM AUTHOR]

Copyright of Journal of Harbin Institute of Technology. Social Sciences Edition / Haerbin Gongye Daxue Xuebao. Shehui Kexue Ban is the property of Harbin Institute of Technology and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

In the era of big data, there exist complex structure between different classes labels. Hierarchical structure, among others, has become a representative one, which is mathematically depicted as a tree-like structure or directed acyclic graph. Most studies in the literature focus on static feature selection in hierarchical information system. In this study, in order to solve the incremental feature selection problem of hierarchical classification in a dynamic environment, we develop two incremental algorithms for this purpose (IHFSGR-1 and IHFSGR-2 for short). As a preliminary step, we propose a new uncertainty measure to quantify the amount of information contained in the hierarchical classification system, and based on this, we develop a non-incremental hierarchical feature selection algorithm. Next, we investigate the updating mechanism of this uncertainty measure upon the arrival of samples, and propose two strategies for adding and deleting features, leading to the development of two incremental algorithms. Finally, we conduct some comparative experiments with several non-incremental algorithms. The experimental results suggest that compared with several non-incremental algorithms, our incremental algorithms can achieve better performance in terms of the classification accuracy and two hierarchical evaluation metrics, and can significantly accelerate the fuzzy rough set-based hierarchical feature selection. [ABSTRACT FROM AUTHOR]

Partial order formal structure analysis (POFSA) is an emerging theory in the field of concept cognitive learning (CCL). Attribute partial order structure diagram (APOSD) is the visual expression of the knowledge structure in POFSA. It has the advantages of explicit expression of the hierarchies of attributes and concise visual expression of the knowledge structure. This paper mainly focuses on the incremental CCL of APOSD in dynamic data circumstances. Firstly, the concept of location information coding of nodes in APOSD is proposed to express the position of nodes in the entire diagram as well as the relationships between nodes, which is an important tool throughout this paper. Secondly, by analyzing the relationship between new objects and objects in the original diagram, dynamic learning strategy for APOSD is proposed. Thirdly, in order to balance the efficiency and accuracy of dynamic learning, a dynamic-static alternating self-learning method for APOSD is proposed, which is an improved incremental learning strategy. Finally, comparative experiments illustrate that compared with non-incremental learning method of APOSD and concept lattice, the two proposed incremental learning methods of APOSD can effectively achieve dynamic self-updating of the knowledge base when processing dynamic data, and provide another perspective for discovering knowledge from the same data. Besides, the effectiveness of the improved incremental learning strategy is verified as well. [ABSTRACT FROM AUTHOR]

Copyright of Journal of Shenzhen University Science & Engineering is the property of Editorial Department of Journal of Shenzhen University Science & Engineering and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Incremental learning is a learning paradigm in which a model is updated continuously as new data becomes available, and its main challenge is to adapt to non-stationary environments without the time-consuming re-training process. Many efforts have been made on incremental supervised learning. However, providing sufficient labeled data remains a major problem. Recently, domain adaptation methods have gained attention. These methods aim to leverage the knowledge from an auxiliary source domain to boost the performance of the model in the target domain by reducing the domain discrepancy between them. Regarding these issues, in the present paper, a proposed model aims to incrementally learn a new domain characterized by drifts due to a non-stationary environment. It utilizes an unsupervised, fuzzy-based domain adaptation to classify data streams faced with concept drift while accounting for a label-agnostic incremental setting in the target domain. Incremental learning updates occur whenever the entropy-based metric indicates uncertainty, ensuring informative samples are integrated. Also, outdated samples are forgotten during the training stage using the dynamic sample weighting strategy. Through experimentation on forty-five tasks, the superiority of the proposed model in handling dynamic adaptation on non-stationary domains is demonstrated, showcasing improvements in accuracy and computational efficiency. [ABSTRACT FROM AUTHOR]

As a widely used data preprocessing method, feature selection with rough sets aims to delete redundant conditional features. However, most of the traditional feature selection methods target to the static data set environment, and the importance of features is used as the base for feature selection. These methods only consider the importance of features themselves and do not consider the impact of features on classification. In order to overcome such shortcomings, we first use the information of knowledge granules to calculate the similarity of samples in the same cluster and samples in different clusters; Secondly, from the perspective of clustering, we stick to the principle that the samples in the same cluster are as close as possible, and the samples in different clusters are as far away as possible, then a feature selection model of knowledge granularity (in short SKG) based on the clustering background is designed; Thirdly, in order to make the SKG model adapt to the reduction of dynamic data sets, we discuss the incremental learning mechanism of sample and feature changes, and two incremental models SKGOA and SKGAA are designed to deal with the dynamic feature reduction when some samples and features are added into the decision system. Finally, some numerical experiments are conducted to assess the performance of the proposed algorithms, and the results shown that our approaches are of a prominent advantage in terms of computational time and classification accuracy. [ABSTRACT FROM AUTHOR]

In this paper, a method is introduced to control the dark knowledge values also known as soft targets, with the purpose of improving the training by knowledge distillation for multi-class classification tasks. Knowledge distillation effectively transfers knowledge from a larger model to a smaller model to achieve efficient, fast, and generalizable performance while retaining much of the original accuracy. The majority of deep neural models used for classification tasks append a SoftMax layer to generate output probabilities and it is usual to take the highest score and consider it the inference of the model, while the rest of the probability values are generally ignored. The focus is on those probabilities as carriers of dark knowledge and our aim is to quantify the relevance of dark knowledge, not heuristically as provided in the literature so far, but with an inductive proof on the SoftMax operational limits. These limits are further pushed by using an incremental decision tree with information gain split. The user can set a desired precision and an accuracy level to obtain a maximal temperature setting for a continual classification process. Moreover, by fitting both the hard targets and the soft targets, one obtains an optimal knowledge distillation effect that mitigates better catastrophic forgetting. The strengths of our method come from the possibility of controlling the amount of distillation transferred non-heuristically and the agnostic application of this model-independent study. [ABSTRACT FROM AUTHOR]

The energy consumption of air-jet looms mainly comes from air compressors. Predicting the air consumption of air-jet looms for the upcoming period is significant for the variable frequency adjustment of air compressors, thereby aiding in energy saving and reducing fabric costs. This paper proposes an innovative method that combines Sliding Time Windows (STW), feature analysis, and incremental learning to improve the accuracy of short-term air consumption prediction. First, the STW method is employed during the data collection phase to enhance data reliability. Through feature analysis, significant factors affecting the air consumption of air-jet looms, beyond traditional research, are explored and incorporated into the prediction model. The experimental results indicate that the introduction of new features improved the model's R2 from 0.905 to 0.950 and reduced the MSE from 32.369 to 16.239. The STW method applied to the same random forest model increased the R2 from 0.906 to 0.950 and decreased the MSE from 32.244 to 16.239. The decision tree method, compared to the linear regression model, improved the R2 from 0.928 to 0.950 and reduced the MSE from 23.541 to 16.239, demonstrating significant predictive performance enhancement. After establishing the optimal model, incremental learning is used to continuously improve the reliability and accuracy of short-term predictions. Experiments show that the incremental learning method, compared to static models, offers better resilience and reliability when new data are collected. The proposed method significantly improves the accuracy of air consumption prediction for air-jet looms, providing strong support for the variable frequency adjustment of air compressors, and contributes to the goals of energy saving and cost reduction. The research results indicate that this method not only enhances prediction accuracy but also provides new insights and methods for future energy-saving research. [ABSTRACT FROM AUTHOR]

Domain Generation Algorithms (DGAs) are algorithms present in most malware used by botnets and advanced persistent threats. These algorithms dynamically generate domain names to maintain and obfuscate communication between the infected device and the attacker's command and control server. Since DGAs are used by many threats, it is extremely important to classify a given DGA according to the threat it is related to. In addition, as new threats emerge daily, classifier models tend to become obsolete over time. Deep neural networks tend to lose their classification ability when retrained with a dataset that is significantly different from the initial one, a phenomenon known as catastrophic forgetting. This work presents a computational scheme composed of a deep learning model based on CNN and natural language processing and an incremental learning technique for class increment through transfer learning to classify 60 DGA families and include a new family to the classifier model, training the model incrementally using some examples from known families, avoiding catastrophic forgetting and maintaining metric levels. The proposed methodology achieved an average precision of 86.75%, an average recall of 83.06%, and an average F1 score of 83.78% with the full dataset, and suffered minimal losses when applying the class increment. [ABSTRACT FROM AUTHOR]

For over a century, the neuro- and pathophysiological, behavioral, and cognitive correlates of consciousness have been an active field of theoretical considerations and empirical research in a wide range of modern disciplines. Conscious cognitive processing of information cannot be observed directly, but might be inferred from step-like discontinuities in learning performance or sudden insight-based improvements in problem solving behavior. It is assumed that a sudden step of knowledge associated with insight requires a creative reorganization of mental representations of task- or problemrelevant information and the restructuration of the task, respectively problem to overcome an cognitive dead-end or impasse. Discontinuities in learning performance or problem solving after an insight event can be used as timetags to capture the time window in which conscious cognitive information processing must have taken place. According to the platform theory of conscious cognitive information processing, the reorganization and restructuration processes, require the maintenance of task- or problem-relevant information in working memory for the operation of executive functions on these mental representations. Electrophysiological evidence suggests that the reorganization and restructuration processes in working memory, that precede insight-based problem solutions are accompanied by an increase in the power of gamma oscillations in cortical areas including the prefrontal cortex. Empirical evidence and theoretical assumptions argue for an involvement of gap junction channels and connexin hemichannels in cortical gamma-oscillations and working memory processes. Discontinuities in learning or problem solving performance might be used as time-tags to investigate the implication of gap junction channels and hemichannels in conscious cognitive processing. [ABSTRACT FROM AUTHOR]

The key point of current research on few shot object detection based on meta learning is how to make better use of the information of support branch to help query branch to recognize novel objects more effectively. However, many current methods fuse the features from support branch and query branch in the depth direction, ignoring the spatial position relationship between features. Therefore, this study proposes non local support attention network. This method not only adds support information into the proposals features, but also fuses the support information with the features that fed into region proposal network. The spatial position relationship between features is considered at the same time. It also adds the information of negative supports to the detection module to help the model distinguish the objects from different categories. This method obtains good performance on base classes and novel classes of COCO2017 dataset, particularly under the case of incremental learning. Compared with method before improvement, 3.3/3.8/4.7 mAP is increased in AP/AP50/AP75 of novel classes. 2.7/0.5/3.3 mAP is increased in AP/AP50/AP75 of the base classes, and outperformed the performance of SOTA (Sort-Of-The Art) model DAnA (Dual-Awareness Attention) under the same setting. [ABSTRACT FROM AUTHOR]

While biological intelligence grows organically as new knowledge is gathered throughout life, Artificial Neural Networks forget catastrophically whenever they face a changing training data distribution. Rehearsal-based Continual Learning (CL) approaches have been established as a versatile and reliable solution to overcome this limitation; however, sudden input disruptions and memory constraints are known to alter the consistency of their predictions. We study this phenomenon by investigating the geometric characteristics of the learner's latent space and find that replayed data points of different classes increasingly mix up, interfering with classification. Hence, we propose a geometric regularizer that enforces weak requirements on the Laplacian spectrum of the latent space, promoting a partitioning behavior. Our proposal, called Continual Spectral Regularizer for Incremental Learning (CaSpeR-IL), can be easily combined with any rehearsal-based CL approach and improves the performance of SOTA methods on standard benchmarks. • We study the geometry of a model's latent space in a Continual Learning setting. • We propose Continual Spectral Regularizer, a geometrically motivated regularizer. • We combine CaSpeR with SOTA rehearsal-based CL approaches in standard benchmarks. • We compare our proposal with recent contrastive-based CL approaches. • We reveal that CaSpeR achieves increased accuracy and reduced forgetting. [ABSTRACT FROM AUTHOR]

In this contribution, the model which is dedicated to reducing the number of false positive instances is proposed. This is a self-tuning model using aggregation functions and time-series data periods. As a case study, the proposed model is tested in the context of phishing link detection. In the proposed model, well-known aggregation functions are applied to combine the confidence values of multiple Classification models for email phishing. The division of the dataset into multiple segments and subsets facilitates the implementation of incremental learning strategies. This approach enables the iterative enhancement of model performance through the training of new data while leveraging previously acquired knowledge. In our research, two datasets are considered, namely the existing PhiUSIIL phishing URL dataset as well as the dataset provided by the FreshMail company are applied. The proposed algorithm achieves a small number of expected false positives. This reduces the costs associated with manual analysis of such cases by domain experts (in the case of incorrect prediction as phishing mail). [ABSTRACT FROM AUTHOR]

The perception system plays a crucial role by integrating LiDAR and various sensors to perform localization and object detection, which ensures the security of intelligent driving. However, existing research indicates that LiDAR is vulnerable to sensor attacks, which lead to inappropriate driving strategies and need effective attack recognition methods. Previous LiDAR attack recognition methods rely on fixed anomaly thresholds obtained from depth map data distributions in specific scenarios as static anomaly boundaries, which lead to reduced accuracy, increased false alarm rates, and a lack of performance stability. To address these problems, we propose an adaptive LiDAR attack recognition framework capable of adjusting to different driving scenarios. This framework initially models the perception system by integrating the vehicle dynamics model and object tracking algorithms to extract data features, subsequently employing Gaussian Processes for the probabilistic modeling of these features. Finally, the framework employs sparsification computing techniques and a sliding window strategy to continuously update the Gaussian Process model with window data, which achieves incremental learning that generates uncertainty estimates as dynamic anomaly boundaries to recognize attacks. The performance of the proposed framework has been evaluated extensively using the real-world KITTI dataset covering four driving scenarios. Compared to previous methods, our framework achieves a 100% accuracy rate and a 0% false positive rate in the localization system, and an average increase of 3.43% in detection accuracy in the detection system across the four scenarios, which demonstrates superior adaptive capabilities. [ABSTRACT FROM AUTHOR]

Last decade demonstrate the massive growth in organizational data which keeps on increasing multi‐fold as millions of records get updated every second. Handling such vast and continuous data is challenging which further opens up many research areas. The continuously flowing data from various sources and in real‐time is termed as streaming data. While deriving valuable statistics from data streams, the variation that occurs in data distribution is called concept drift. These drifts play a significant role in a variety of disciplines, including data mining, machine learning, ubiquitous knowledge discovery, quantitative decision theory, and so forth. As a result, a substantial amount of research is carried out for studying methodologies and approaches for dealing with drifts. However, the available material is scattered and lacks guidelines for selecting an effective technique for a particular application. The primary novel objective of this survey is to present an understanding of concept drift challenges and allied studies. Further, it assists researchers from diverse domains to accommodate detection and adaptation algorithms for concept drifts in their applications. Overall, this study aims to contribute to deeper insights into the classification of various types of drifts and methods for detection and adaptation along with their key features and limitations. Furthermore, this study also highlights performance metrics used to evaluate the concept drift detection methods for streaming data. This paper presents the future research scope by highlighting gaps in the existing literature for the development of techniques to handle concept drifts. This article is categorized under:Algorithmic Development > Ensemble MethodsApplication Areas > Data Mining Software ToolsFundamental Concepts of Data and Knowledge > Big Data Mining [ABSTRACT FROM AUTHOR]

The article provides a comprehensive overview of incremental learning in the context of big data processing. The basic concepts, modern approaches, and key aspects of incremental learning are considered. The advantages of this approach for processing large amounts of data are analyzed, including the efficient use of computing resources, the ability to process streaming data in real time, and adaptability to changes in data. The main limitations and challenges, such as the problem of "catastrophic forgetting", the difficulty of balancing new and old knowledge, dependence on the order of data arrival, and potential loss of accuracy, are investigated. An analysis of specific problems is presented, including the handling of conceptual drift, unbalanced classes, and missing features. Applications of incremental learning in various fields, including data analytics, robotics, autonomous driving, and activity recognition, are discussed. We suggest directions for future research to address the identified problems and improve the effectiveness of incremental learning in the context of big data. [ABSTRACT FROM AUTHOR]

Copyright of Journal of Frontiers of Computer Science & Technology is the property of Beijing Journal of Computer Engineering & Applications Journal Co Ltd. and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Copyright of Journal of Jilin University (Science Edition) / Jilin Daxue Xuebao (Lixue Ban) is the property of Zhongguo Xue shu qi Kan (Guang Pan Ban) Dian zi Za zhi She and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Artificial intelligence applications could be more powerful and comprehensive by incorporating the ability of inference, which could be achieved by probabilistic inference over implicit relations. It is significant yet challenging to represent implicit relations among observed variables and latent ones like disease etiologies and user preferences. In this paper, we propose the BN with multiple latent variables (MLBN) as the framework for representing the dependence relations, where multiple latent variables are incorporated to describe multi-dimensional abstract concepts. However, the efficiency of MLBN learning and effectiveness of MLBN based applications are still nontrivial due to the presence of multiple latent variables. To this end, we first propose the constraint induced and Spark based algorithm for MLBN learning, as well as several optimization strategies. Moreover, we present the concept of variation degree and further design a subgraph based algorithm for incremental learning of MLBN. Experimental results suggest that our proposed MLBN model could represent the dependence relations correctly. Our proposed method outperforms some state-of-the-art competitors for personalized recommendation, and facilitates some typical approaches to achieve better performance. [ABSTRACT FROM AUTHOR]

We present the first major release of the Classification Algorithms Comparison Pipeline (CACP). The proposed software enables one to compare newly developed classification algorithms in Python with other classifiers to evaluate classification performance and ensure both outcomes’ reproducibility and statistical reliability. CACP simplifies and accelerates the entire classifier evaluation process considerably and helps prepare the professional documentation of the experiments conducted. The upgrade introduces enhancements to existing tools and adds new features: (1) - support for River machine learning library datasets in incremental learning, (2) - capability to include user-defined datasets, (3) - use of River classifiers for incremental learning, (4) - use of River metrics for incremental learning, (5) - flexibility to create user-defined metrics, (6) - record-by-record testing for incremental learning, (7) - enhanced summary of incremental testing results with dynamic visualization of the learning process, (8) - Graphical User Interface (GUI).

The weld crack leakage due to stress concentration and external load is a significant safety risk in pressure pipelines. Microstructural variations and dynamic propagation lead to unpredictable changes in leakage rate over time and conditions. To address the above problems, a novel framework called OILS-TCN for weld crack pattern recognition and leakage rate prediction is proposed. Firstly, the adaptive threshold optimization algorithm is introduced into the self-organizing incremental neural network to update and increase the crack leakage pattern. Secondly, the depth first search algorithm is combined with the radial basis function neural network to perform online increment labelling of the leakage state. Then, according to the attenuation characteristics of acoustic emission signals, a portable input-attention module is designed to add different weights to the input sequence. Finally, the accurate prediction of leakage rate under different conditions is realized based on the temporal convolutional network. Compared with other advanced methods, the proposed method has obvious advantages in the adaptability and accuracy of pipeline weld crack leakage rate prediction. In addition, the validity and necessity of each part of the framework proposed are discussed based on ablation experiments. The proposed method can predict the leakage rate in real time without modifying the hyperparameters of the model, and can provide a powerful guide for the online monitoring of the leakage AE technology of pressure pipeline in complex systems.

Collaborative robots (cobots) are increasingly integrated into Industry 4.0 dynamic manufacturing environments that require frequent system reconfiguration due to changes in cobot paths and payloads. This necessitates fast methods for identifying payload inertial parameters to compensate the cobot controller and ensure precise and safe operation. Our prior work used Incremental Ensemble Model (IEM) to identify payload parameters, eliminating the need for an excitation path and thus removing the separate identification step. However, this approach suffers from catastrophic forgetting. This paper introduces a novel incremental ensemble learning method that addresses the problem of catastrophic forgetting by adding a new weak learner to the ensemble model for each new training bag. Moreover, it proposes a new classification model that assists the ensemble model in identifying which weak learner provides the most accurate estimation for new input data. The proposed method incrementally updates the identification model while the cobot navigates any task path, maintaining accuracy on old weak learner even after updating with new data. Validation performed on the Franka Emika cobot showcases the model’s superior accuracy and adaptability, effectively eliminating the problem of catastrophic forgetting.

Natural-language dialog is key for an intuitive human–robot interaction. It can be used not only to express humans’ intents but also to communicate instructions for improvement if a robot does not understand a command correctly. It is of great importance to let robots learn from such interaction experiences in an incremental way to allow them to improve their behaviors or avoid mistakes in the future. In this paper, we propose a system to achieve such incremental learning of complex high-level behavior from natural interaction and demonstrate its implementation on a humanoid robot. Our system deploys large language models (LLMs) for high-level orchestration of the robot’s behavior based on the idea of enabling the LLM to generate Python statements in an interactive console to invoke both robot perception and action. Human instructions, environment observations, and execution results are fed back to the LLM, thus informing the generation of the next statement. Since an LLM can misunderstand (potentially ambiguous) user instructions, we introduce incremental learning from the interaction, which enables the system to learn from its mistakes. For that purpose, the LLM can call another LLM responsible for code-level improvements in the current interaction based on human feedback. Subsequently, we store the improved interaction in the robot’s memory so that it can later be retrieved on semantically similar requests. We integrate the system in the robot cognitive architecture of the humanoid robot ARMAR-6 and evaluate our methods both quantitatively (in simulation) and qualitatively (in simulation and real-world) by demonstrating generalized incrementally learned knowledge.

The stock market’s volatility, noise, and information overload necessitate efficient prediction methods. Forecasting index prices in this environment is complex due to the non-linear and non-stationary nature of time series data generated from the stock market. Machine learning and deep learning have emerged as powerful tools for identifying financial data patterns and generating predictions based on historical trends. However, updating these models in real-time is crucial for accurate predictions. Deep learning models require extensive computational resources and careful hyperparameter optimization, while incremental learning models struggle to balance stability and adaptability. This paper proposes a novel hybrid bidirectional-LSTM (H.BLSTM) model that combines incremental learning and deep learning techniques for real-time index price prediction, addressing these scalability and memory challenges. The method utilizes both univariate time series derived from historical index prices and multivariate time series incorporating technical indicators. Implementation within a real-time trading system demonstrates the method’s effectiveness in achieving more accurate price forecasts for major stock indices globally through extensive experimentation. The proposed model achieved an average mean absolute percentage error of 0.001 across nine stock indices, significantly outperforming traditional models. It has an average forecasting delay of 2 s, making it suitable for real-time trading applications.