Chao Xue, Yao Wang, Mengqiao Liu, Di Liang, Xingsheng Han, Peiyang Liu, Xianjie Wu, Chenyao Lu, Lei Jiang, Yu Lu, Haibo Shi, Shuang Liang, Minlong Peng, Flora D. Salim. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Compared with the optimization technology of retraining models, fine-tuning pretrained neural networks have been widely used in industrial process monitoring because of their high precision and low training cost. However, data privacy protection in industry makes it impossible to improve the upstream pertaining model to avoid negative transfer when the target domain and the source domain are too different or the pertaining model is maliciously attacked. To solve the above problem, we proposed a fine-tuning method assisted by target domain knowledge for industrial time-lag parameter prediction. This method uses an Auto-Encoder to learn the representation of black-box and time-lag knowledge in the target domain. Then, the black-box and time-lag knowledge are used as auxiliary information to update the high-level weights of the pretrained network. At the same time, we proposed an auxiliary learning method that can dynamically update weights without introducing new parameters and provide training methods for different neural network optimizers. The experimental results on the heating furnace temperature prediction and wind condition prediction of wind farms demonstrate that the prediction performance can be effectively improved, and the defective inheritance of the pretrained model can be effectively reduced to avoid negative transfer.Note to Practitioners—Due to the complexity of industrial processes, the detection of key parameters is time-consuming and it is difficult to obtain labeled samples. With the widespread use of large models in computer vision and natural language processing, fine-tuning pretrained models can alleviate these challenges. However, the problem that comes with it is safety. The industrial prediction model cannot be built when the publicly available pretrained model is attacked. Moreover, the pretrained model could not be modified because the pretrained data could not be obtained. Therefore, this paper used the time-lag prior knowledge of industrial data to modify the pretrained model in the fine-tuning stage, and helps improve the pretrained accuracy and model robustness by mining the knowledge of existing data. Two groups of experiments show that the auxiliary knowledge-based fine-tuning mechanism can significantly improve the prediction accuracy and alleviate the defect inheritance of the pretrained model, which is valuable for the application of large models in the field of industrial parameter prediction.

Deep learning models have gained significant attention and application in recent years to improve the accuracy and efficiency of industrial time series prediction. However, the dynamic changes in industrial processes present a key challenge for data-driven models. Specifically, the performance of deployed models deteriorates over time and fails to adapt to new operating conditions. Currently, two common update methods exist: Retraining the model using historical and new operating data, which incurs high computation and storage costs, or incrementally fine-tuning the model solely using new data, which leads to catastrophic forgetting of learned patterns. To address these issues, this article proposes an adaptive continual learning method for nonstationary industrial time series prediction. Our approach tackles the problems by hint-based network parameter learning to retain the dark knowledge from previous tasks and avoid catastrophic forgetting of accumulated knowledge. In addition, we design a soft buffer to aid memory and learning of key patterns under the current operating condition. Lastly, a time-sensitive activation function is proposed to endow the neural network with time-evolving properties, thereby enhancing the model's generalization ability. Compared with other update methods and different continual learning methods, the superiority of our method is validated on solar power generation data and real data of grinding and grading process.

Pancreatic cancer, characterized by its notable prevalence and mortality rates, demands accurate lesion delineation for effective diagnosis and therapeutic interventions. The generalizability of extant methods is frequently compromised due to the pronounced variability in imaging and the heterogeneous characteristics of pancreatic lesions, which may mimic normal tissues and exhibit significant inter-patient variability. Thus, we propose a generalization framework that synergizes pixel-level classification and regression tasks, to accurately delineate lesions and improve model stability. This framework not only seeks to align segmentation contours with actual lesions but also uses regression to elucidate spatial relationships between diseased and normal tissues, thereby improving tumor localization and morphological characterization. Enhanced by the reciprocal transformation of task outputs, our approach integrates additional regression supervision within the segmentation context, bolstering the model's generalization ability from a dual-task perspective. Besides, dual self-supervised learning in feature spaces and output spaces augments the model's representational capability and stability across different imaging views. Experiments on 594 samples composed of three datasets with significant imaging differences demonstrate that our generalized pancreas segmentation results comparable to mainstream in-domain validation performance (Dice: 84.07%). More importantly, it successfully improves the results of the highly challenging cross-lesion generalized pancreatic cancer segmentation task by 9.51%. Thus, our model constitutes a resilient and efficient foundational technological support for pancreatic disease management and wider medical applications. The codes will be released at https://github.com/SJTUBME-QianLab/Dual-Task-Seg.

Yougang Lyu, Lingyong Yan, Shuaiqiang Wang, Haibo Shi, Dawei Yin, Pengjie Ren, Zhumin Chen, Maarten de Rijke, Zhaochun Ren. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Large language models (LLMs) are proven to benefit a lot from retrieval-augmented generation (RAG) in alleviating hallucinations confronted with knowledge-intensive questions.RAG adopts information retrieval techniques to inject external knowledge from semanticrelevant documents as input contexts.However, since today's Internet is flooded with numerous noisy and fabricating content, it is inevitable that RAG systems are vulnerable to these noises and prone to respond incorrectly.To this end, we propose to optimize the retrieval-augmented GENERATOR with an Adversarial Tuning Multi-agent system (ATM).The ATM steers the GENERATOR to have a robust perspective of useful documents for question answering with the help of an auxiliary ATTACKER agent through adversarially tuning the agents for several iterations.After rounds of multi-agent iterative tuning, the GENERA-TOR can eventually better discriminate useful documents amongst fabrications.The experimental results verify the effectiveness of ATM and we also observe that the GENERATOR can achieve better performance compared to the state-of-the-art baselines.The code is available at https://github.com/chuhac/ATM-RAG.

Visual question answering (VQA) has been intensively studied as a multimodal task, requiring efforts to bridge vision and language for correct answer inference. Recent attempts have developed various attention-based modules for solving VQA tasks. However, the performance of model inference is largely bottlenecked by visual semantic comprehension. Most existing detection methods rely on bounding boxes, remaining a serious challenge for VQA models to comprehend and correctly infer the causal nexus of contextual object semantics in images. To this end, we propose a finer model framework without bounding boxes in this work, termedLooking Out of Instance Semantics (LOIS)to address this crucial issue. LOIS can achieve more fine-grained feature descriptions to generate visual facts. Furthermore, to overcome the label ambiguity caused by instance masks, two types of relation attention modules: 1) intra-modality and 2) inter-modality, are devised to infer the correct answers from different visual features. Specifically, we implement a mutual relation attention module to model sophisticated and deeper visual semantic relations between instance objects and background information. In addition, our proposed attention model can further analyze salient image regions by focusing on important word-related questions. Experimental results on four benchmark VQA datasets prove that our proposed method has favorable performance in improving visual reasoning capability.

Industrial time series, as a kind of data that responds to production process information, can be analyzed and predicted for effective monitoring of industrial production processes. There are problems of data shortage and algorithm cold start in industrial modeling process caused by complex working conditions, change of data acquisition environment, and short running time of equipment. As a result, the accuracy of the existing data-driven industrial time series prediction algorithm is greatly limited. To address the aforementioned problems, we propose a new time series prediction method for industrial processes under limited data based on dynamic transfer learning in this work. This method aims to effectively use historical data of similar equipment or working conditions rather than discard them to help establish an industrial time series prediction model with limited target data. In this method, first, historical data are divided into multiple batches, and then a new multisource transfer learning framework with dynamic maximum mean difference loss is established according to the distribution distance between each batch of historical data and the limited target data at the current moment. The framework also combines multitask learning methods to establish multistep prediction model for online learning in industrial processes. Compared with other commonly used methods, experiments on two real-world datasets of solar power generation prediction and heating furnace temperature prediction demonstrate the effectiveness of the proposed method.

In recent years, spiking neural networks (SNNs), which originated from the theoretical basis of neuroscience, have attracted neuromorphic computing and brain-like computing due to their advantages, such as neural dynamics and coding mechanism, which are similar to biological neurons. SNNs have become one of the mainstream frameworks in the field of brain-like computing. However, most of the Leaky Integrate-and-Fire (LIF) neuron models currently used by SNNs based on direct training of backpropagation (BP) do not consider the changes in the recurrent connections and the dynamic strength of neuron connections over time. This study presented the LIF neuron model with recurrent connections and a method for dynamically changing the presynaptic currents. Recurrent LIF neurons have an additional cyclic connection compared with classic LIF neurons. Their postsynaptic current stimulates a change in membrane potential at the next time point. Their dynamics were more similar to the activities of biological neurons. We also proposed an efficient and flexible BP training method for recurrent LIF neurons. On the basis of the above methods, we proposed the recurrent SNN with dynamic presynaptic currents based on backpropagation (RDS-BP). We test the proposed RDS-BP on three image data sets (MNIST, Fashion-MNIST and CIFAR-10) and two text data sets (IMDB and TREC). The results showed that the performance of RDS-BP not only exceeded the naive SNN models based on BP but also exceeded the SNN methods proposed in previous studies in recent years, which had excellent performance in previous experiments. Our work provides a new LIF neuron model with a recurrent connection and dynamic presynaptic current and a BP training arrangement for the proposed neuron, which could merit developments with neuromorphic and brain-like computing.

Model-based reinforcement learning (MBRL) approaches have demonstrated great potential in handling complex tasks with high sample efficiency. However, MBRL struggles with asymptotic performance compared to model-free reinforcement learning (MFRL). In this paper, we present a long-horizon policy optimization method, namely model-based deterministic policy gradient (MBDPG), for efficient exploitation of the learned dynamics model through multi-step gradient information. First, we approximate the dynamics of the environment with a parameterized linear combination of an ensemble of Gaussian distributions. Moreover, the dynamics model is equipped with a memory module and trained on a multi-step prediction task to reduce cumulative error. Second, successful experience is used to guide the policy at the early stage of training to avoid ineffective exploration. Third, a clipped double value network is expanded in the learned dynamics to reduce overestimation bias. Finally, we present a deterministic policy gradient approach in the model that backpropagates multi-step gradient along the imagined trajectories. Our method shows higher sampling efficiency than the state-of-the-art MFRL methods while maintaining better convergence performance and time efficiency compared to the SOAT MBRL.

Planning in learned dynamics models has proven to have great potential in improving sample efficiency. However, learning the latent representation that embeds to model dynamics from high-dimensional observation is still challenging. We here propose a model-based policy gradient method, that quickly learns an optimal policy directly from pixel frames. First, the dynamics model is learned with a SENet architecture, that explicitly incorporates attention and gating mechanism to differentiate features for the latent representation. Second, the policy in the early stage is guided with successful experience to extract the intention of the experts thus to speed up the convergence. Finally, we use model rollout to decrease the value estimation bias and multi-step policy gradients to update the policy. Our approach outperforms state-of-the-art algorithms on multiple benchmark tasks in sampling efficiency and convergence performance.

Unsupervised Image-to-Image generation has obtained many studies recently, where generative adversarial networks (GANs) have developed as an effective model. Cycle consistency or dual learning guides the GAN from aligned image pairs to the unpaired training set and can be applied for many applications based on the image-to-image generation. However, for many tasks, error accumulation, which can be produced in the progress of image reconstruction, affects the realism and quality of the generated images. To eliminate error accumulation and guide the adversarial learning, in this paper, we propose dual-cycle consistent GAN (DucGAN), a novel approach for cross-domain image-to-image generation. The key idea is to introduce dual-cycle learning, which restrains error accumulation. In our method, inter-cycle and extra-cycle are based on the cycle consistency, while the output from intercycle is cast as a new augmented input in extra-cycle. On the one hand, reconstruction loss in extra-cycle can constrain and guide the training. On the other hand, dual-cycle learning is self-augmented. Our extensive experiments on image-to-image generation tasks show that DucGAN is effective and superior to several state-of-the-art GANs.

The performance of many computer vision and machine learning algorithms are heavily depend on the distance metric between samples. It is necessary to exploit abundant of side information like pairwise constraints to learn a robust and reliable distance metric[2, 3]. Let D = {(xl i ,xj,yi j)} l i, j=1 denotes the labeled training set for the target task, wherein xi, x j ∈ Rd and yi j = ±1 indicates xl i and xl i are similar/dissimilar to each other. Then, a metric is usually learned to minimize the distance between the data from the same class and maximize their distance otherwise. This leads to the following loss function for learning the metric A:

Quality prediction capability plays a crucial role in ensuring high quality of products as well as reliability of systems during manufacturing processes. In this study, the prediction reasoning mechanism based on fuzzy Petri nets (FPNs) is presented. The feasibility and validity of the reasoning mechanism are proved. A quality prediction model and reasoning process is presented for submerged arc welding process. This application example illustrates the proposed method based on FPNs predict the potential quality problems and quality accidents effectively.

This paper presents a backward reasoning approach based on fuzzy Petri nets (FPNs). This approach takes full advantage of the structural and behavioral properties of a FPN. It can identify middle places by a vector-computational manner rather than the conventional search way, improving inference efficiency. To reduce the complexity and scale of a FPN, man-machine interaction is introduced to it. It is suggested that high efficiency and low costs which an inference method brings in practice, play a more important role than the operational efficiency of the method itself. An instance illustrates that the reasoning approach is feasible and effective.