Domain-specific Representations Research Articles

Order shaped by cognition. Evidence for (and against) the effect of domain-general biases on word and morpheme order

Abstract A foundational goal of linguistics has been to understand why languages look the way they do. A range of possible explanations exist – from domain-specific representations to cognition-external factors like history and grammaticalization – and all of these undoubtedly play some role. But determining exactly how these link with specific features of language remains challenging, and the role of domain-specific mechanisms has been particularly contentious. In this paper, I highlight a growing new approach, which uses artificial language experiments to link individual-level biases to cross-linguistic trends in language structure. Using word and morpheme order as case studies, I will show how a range of different paradigms and learner populations allow us to make progress on this crucial issue in linguistics. I will focus on typological trends in word and morpheme order. For some ordering trends, experimental evidence points to variation across populations, suggesting that the best explanation for these patterns likely has its root in language history and grammaticalization. In other cases, the evidence points to the role of universal but domain-general cognitive biases, like transparency and simplicity. These domain-general biases interact with linguistic representations in important ways. Taken together these studies help adjudicate between alternative explanations for a number of specific ordering patterns and suggest a new sense of domain-specificity in the evolution of language.

A Domain Generation Diagnosis Framework for Unseen Conditions Based on Adaptive Feature Fusion and Augmentation

Emerging deep learning-based fault diagnosis methods have advanced in the current industrial scenarios of various working conditions. However, the prerequisite of obtaining target data in advance limits the application of these models to practical engineering scenarios. To address the challenge of fault diagnosis under unseen working conditions, a domain generation framework for unseen conditions fault diagnosis is proposed, which consists of an Adaptive Feature Fusion Domain Generation Network (AFFN) and a Mix-up Augmentation Method (MAM) for both the data and domain spaces. AFFN is utilized to fuse domain-invariant and domain-specific representations to improve the model’s generalization performance. MAM enhances the model’s exploration ability for unseen domain boundaries. The diagnostic framework with AFFN and MAM can effectively learn more discriminative features from multiple source domains to perform different generalization tasks for unseen working loads and machines. The feasibility of the proposed unseen conditions diagnostic framework is validated on the SDUST and PU datasets and achieved peak diagnostic accuracies of 94.15% and 93.27%, respectively.

Towards diverse image-to-image translation via adaptive normalization layer and contrast learning

A nice image-to-image translation framework is able to acquire an explicit and credible mapping relationship between the source domain and target domains while satisfying two requirements. One is simplicity, the other is extensibility over multiple translation tasks. To this end, we design a concise but versatile generative model for image-to-image translation. Our method includes three major ingredients. First, inspired by popular unconditional normalization layers, named Spatially Adaptive Normalization(SPADE). We introduce a novel Semantics-Appearance Spatially Adaptive Normalization (SA-SPADE), taking into account both semantic structure and style appearance. This enables semantic composition and style appearance information to be sufficiently captured and integrated by our normalization layers. Thanks to SA-SPADE, our model extends to multiple image-to-image translation tasks in an unsupervised or supervised way. Second, we carefully designed two symmetrical network branches to provide semantic and appearance information for our normalization layer, namely Semantic Branch (SB) and Appearance Branch(AB) respectively. Third, we propose novel Semantic-aware Contrastive Loss (SCL) and Appearance-aware Contrastive Loss (ACL)based on newly un-/self- supervised contrastive learning. That is, SCL guarantees domain-invariant (e.g., pose, structure) representations between the generated image and the input image, while ACL ensures domain-specific representations (e.g., color, texture) between the generated image and the reference image. As a result, we verify the effectiveness of our method by comparing it with various task-dependent image translation models in both qualitative and quantitative evaluations.

Forge: A Tool and Language for Teaching Formal Methods

This paper presents the design of Forge , a tool for teaching formal methods gradually. Forge is based on the widely-used Alloy language and analysis tool, but contains numerous improvements based on more than a decade of experience teaching Alloy to students. Although our focus has been on the classroom, many of the ideas in Forge likely also apply to training in industry. Forge offers a progression of languages that improve the learning experience by only gradually increasing in expressive power. Forge supports custom visualization of its outputs, enabling the use of widely-understood domain-specific representations. Finally, Forge provides a variety of testing features to ease the transition from programming to formal modeling. We present the motivation for and design of these aspects of Forge, and then provide a substantial evaluation based on multiple years of classroom use.

Why comparing matters – on case comparisons in organic chemistry

When working with domain-specific representations such as structural molecular representations and reaction mechanisms, learners need to be engaged in multiple cognitive operations, from attending to relevant areas of representations, linking implicit information to structural features, and making meaningful connections between information and reaction processes. For these processes, appropriate instruction, such as a clever task design, becomes a crucial factor for successful learning. Chemistry learning, and especially organic chemistry, merely addressed meaningful task design in classes, often using more reproduction-oriented predict-the-product tasks. In recent years, rethinking task design has become a major focus for instructional design in chemistry education research. Thus, this perspective aims to illustrate the theoretical underpinning of comparing cases from different perspectives, such as the structure-mapping theory, the cognitive load theory, and the variation theory, and outlines, based on the cognitive theory of multimedia learning, how instructors can support their students. Variations of this task design in the chemistry classroom and recommendations for teaching with case comparisons based on current state-of-the-art evidence from research studies in chemistry education research are provided.

Triple Feature Disentanglement for One-Stage Adaptive Object Detection

In recent advancements concerning Domain Adaptive Object Detection (DAOD), unsupervised domain adaptation techniques have proven instrumental. These methods enable enhanced detection capabilities within unlabeled target domains by mitigating distribution differences between source and target domains. A subset of DAOD methods employs disentangled learning to segregate Domain-Specific Representations (DSR) and Domain-Invariant Representations (DIR), with ultimate predictions relying on the latter. Current practices in disentanglement, however, often lead to DIR containing residual domain-specific information. To address this, we introduce the Multi-level Disentanglement Module (MDM) that progressively disentangles DIR, enhancing comprehensive disentanglement. Additionally, our proposed Cyclic Disentanglement Module (CDM) facilitates DSR separation. To refine the process further, we employ the Categorical Features Disentanglement Module (CFDM) to isolate DIR and DSR, coupled with category alignment across scales for improved source-target domain alignment. Given its practical suitability, our model is constructed upon the foundational framework of the Single Shot MultiBox Detector (SSD), which is a one-stage object detection approach. Experimental validation highlights the effectiveness of our method, demonstrating its state-of-the-art performance across three benchmark datasets.

Exploring One-Shot Semi-supervised Federated Learning with Pre-trained Diffusion Models

Recently, semi-supervised federated learning (semi-FL) has been proposed to handle the commonly seen real-world scenarios with labeled data on the server and unlabeled data on the clients. However, existing methods face several challenges such as communication costs, data heterogeneity, and training pressure on client devices. To address these challenges, we introduce the powerful diffusion models (DM) into semi-FL and propose FedDISC, a Federated Diffusion-Inspired Semi-supervised Co-training method. Specifically, we first extract prototypes of the labeled server data and use these prototypes to predict pseudo-labels of the client data. For each category, we compute the cluster centroids and domain-specific representations to signify the semantic and stylistic information of their distributions. After adding noise, these representations are sent back to the server, which uses the pre-trained DM to generate synthetic datasets complying with the client distributions and train a global model on it. With the assistance of vast knowledge within DM, the synthetic datasets have comparable quality and diversity to the client images, subsequently enabling the training of global models that achieve performance equivalent to or even surpassing the ceiling of supervised centralized training. FedDISC works within one communication round, does not require any local training, and involves very minimal information uploading, greatly enhancing its practicality. Extensive experiments on three large-scale datasets demonstrate that FedDISC effectively addresses the semi-FL problem on non-IID clients and outperforms the compared SOTA methods. Sufficient visualization experiments also illustrate that the synthetic dataset generated by FedDISC exhibits comparable diversity and quality to the original client dataset, with a neglectable possibility of leaking privacy-sensitive information of the clients.

Industrial process fault diagnosis based on feature enhanced meta-learning toward domain generalization scenarios

With the rise of the Internet of Things, computing science, and artificial intelligence, data-driven fault diagnosis methods for industrial processes have attracted considerable attention. However, these methods typically assume that the distribution of the test data (target domains) is similar to that of the training data (source domains), making it difficult to deploy them in actual scenarios because the distribution of the collected data varies with operating conditions. Faced with this challenge, a novel fault diagnosis method based on feature enhanced meta-learning (FEML) is proposed to learn domain generalization knowledge effectively. It aims to jointly extract domain-invariant information that can be transferred across domains, and domain-specific information related to classification labels. First, feature extractors based on gated recurrent units (GRUs) are designed to capture domain-invariant and domain-specific representations with time-sequence characteristics from industrial process data. Second, a meta-learning strategy is introduced to enhance the domain-specific representation, thus improving the generalization ability of the model without accessing the target domains. Finally, the Tennessee Eastman process (TEP), Fed-Batch Fermentation Penicillin process (FBFP), and Three-Phase Flow process (TPFP) are used to verify the effectiveness of FEML. The results illustrate that the FEML is robust and outperforms other related methods. The diagnostic accuracies for TEP, FBFP, and TPFP are 87.11%, 95.94%, and 83.07%, respectively.

Influence of sport expertise in facilitating and inhibiting the recognition of the opponent's intentions in sailing.

Starting from the proposed role of the mirror neuron system in the recognition of the intention underlying the actions of others, an experimental paradigm was implemented to test the role of sailing motor expertise in predicting the outcome of a competitor's action. It was hypothesized that subjects with experience in sailing would correctly interpret the maneuver performed due to the activation of domain specific motor representations of the same movements and that subjects who practiced a sport different from sailing would perform worse because of the activation of irrelevant motor patterns. For doing so, a series of video clips, in which a professional sailor performed a tack or a feint, have been manipulated so that the video clips would stop at the moment of the dunkin, namely, when the boat acquires speed to tack or continue straight ahead. The task consisted in predicting whether the action following the dunkin was an actual tack or a feint. The performance of 87 subjects, divided into three subgroups (sailors, tennis players, sedentary), was evaluated in terms of accuracy in identifying the sailor's intentions and correlated to age, gender, manual dominance, education, job, hours spent weekly playing videogames, and experience in playing sports. Results showed that the percentage of correct identifications of the intention to do a tack or feint was the highest in the group of sailors and the lowest in tennis players. An inverse relation between tennis experience and ability in recognizing the sailor's intention was found in the group of tennis players. Gender, age, manual dominance, education, job, and experience with videogames were not found to be correlated with performance. Findings support the possible implication of the mirror neuron system in maneuver detection in sailing and may be a starting point for the development of psychological training in this sport.

Climate Change Sentiment Analysis using Domain Specific Bidirectional Encoder Representations from Transformers

Climate Change Sentiment Analysis using Domain Specific Bidirectional Encoder Representations from Transformers

A Cross-Scale Transformer and Triple-View Attention Based Domain-Rectified Transfer Learning for EEG Classification in RSVP Tasks.

Rapid serial visual presentation (RSVP)-based brain-computer interface (BCI) is a promising target detection technique by using electroencephalogram (EEG) signals. However, existing deep learning approaches seldom considered dependencies of multi-scale temporal features and discriminative multi-view spectral features simultaneously, which limits the representation learning ability of the model and undermine the EEG classification performance. In addition, recent transfer learning-based methods generally failed to obtain transferable cross-subject invariant representations and commonly ignore the individual-specific information, leading to the poor cross-subject transfer performance. In response to these limitations, we propose a cross-scale Transformer and triple-view attention based domain-rectified transfer learning (CST-TVA-DRTL) for the RSVP classification. Specially, we first develop a cross-scale Transformer (CST) to extract multi-scale temporal features and exploit the dependencies of different scales features. Then, a triple-view attention (TVA) is designed to capture spectral features from triple views of multi-channel time-frequency images. Finally, a domain-rectified transfer learning (DRTL) framework is proposed to simultaneously obtain transferable domain-invariant representations and untransferable domain-specific representations, then utilize domain-specific information to rectify domain-invariant representations to adapt to target data. Experimental results on two public RSVP datasets suggests that our CST-TVA-DRTL outperforms the state-of-the-art methods in the RSVP classification task. The source code of our model is publicly available in https://github.com/ljbuaa/CST_TVA_DRTL.

Single-/Multi-Source Domain Adaptation via domain separation: A simple but effective method

Domain adaptation (DA) aims to reduce knowledge gap between domains and improve the prediction ability of models in the target domain. However, the representations learned from feature extraction network often contain redundant information, which is harmful to domain alignment. In addition, many methods only focus on either single-source DA task or multi-source DA task, which limits their real-world applications. In this paper, we propose a simple but effective method called Category and Domain Features Augmentation (CDFA), which consists of two components: Contrastive Classifier Network (CCN) and Domain-specific Learning Network (DSLN). CDFA can remove the specific representation at the feature extraction stage to alleviate transfer difficulty, where CCN is used to increase the probabilistic output of samples and avoid misclassification, and DSLN facilitates the separation of redundant information from all representations by learning domain-specific representations. Empirical evaluations on several cross-domain benchmarks under single-source and multi-source DA scenarios illustrate the competitive performance of CDFA with respect to the state-of-the-art.

POEM: Polarization of Embeddings for Domain-Invariant Representations

Handling out-of-distribution samples is a long-lasting challenge for deep visual models. In particular, domain generalization (DG) is one of the most relevant tasks that aims to train a model with a generalization capability on novel domains. Most existing DG approaches share the same philosophy to minimize the discrepancy between domains by finding the domain-invariant representations. On the contrary, our proposed method called POEM acquires a strong DG capability by learning domain-invariant and domain-specific representations and polarizing them. Specifically, POEM co-trains category-classifying and domain-classifying embeddings while regularizing them to be orthogonal via minimizing the cosine-similarity between their features, i.e., the polarization of embeddings. The clear separation of embeddings suppresses domain-specific features in the domain-invariant embeddings. The concept of POEM shows a unique direction to enhance the domain robustness of representations that brings considerable and consistent performance gains when combined with existing DG methods. Extensive simulation results in popular DG benchmarks with the PACS, VLCS, OfficeHome, TerraInc, and DomainNet datasets show that POEM indeed facilitates the category-classifying embedding to be more domain-invariant.

Domain Generalization for Activity Recognition via Adaptive Feature Fusion

Human activity recognition requires the efforts to build a generalizable model using the training datasets with the hope to achieve good performance in test datasets. However, in real applications, the training and testing datasets may have totally different distributions due to various reasons such as different body shapes, acting styles, and habits, damaging the model’s generalization performance. While such a distribution gap can be reduced by existing domain adaptation approaches, they typically assume that the test data can be accessed in the training stage, which is not realistic. In this article, we consider a more practical and challenging scenario: domain-generalized activity recognition (DGAR) where the test dataset cannot be accessed during training. To this end, we propose Adaptive Feature Fusion for Activity Recognition (AFFAR) , a domain generalization approach that learns to fuse the domain-invariant and domain-specific representations to improve the model’s generalization performance. AFFAR takes the best of both worlds where domain-invariant representations enhance the transferability across domains and domain-specific representations leverage the model discrimination power from each domain. Extensive experiments on three public HAR datasets show its effectiveness. Furthermore, we apply AFFAR to a real application, i.e., the diagnosis of Children’s Attention Deficit Hyperactivity Disorder (ADHD), which also demonstrates the superiority of our approach.

Rethinking adversarial domain adaptation: Orthogonal decomposition for unsupervised domain adaptation in medical image segmentation.

Medical image segmentation methods based on deep learning have made remarkable progress. However, such existing methods are sensitive to data distribution. Therefore, slight domain shifts will cause a decline of performance in practical applications. To relieve this problem, many domain adaptation methods learn domain-invariant representations by alignment or adversarial training whereas ignoring domain-specific representations. In response to this issue, this paper rethinks the traditional domain adaptation framework and proposes a novel orthogonal decomposition adversarial domain adaptation (ODADA) architecture for medical image segmentation. The main idea behind our proposed ODADA model is to decompose the input features into domain-invariant and domain-specific representations and then use the newly designed orthogonal loss function to encourage their independence. Furthermore, we propose a two-step optimization strategy to extract domain-invariant representations by separating domain-specific representations, fighting the performance degradation caused by domain shifts. Encouragingly, the proposed ODADA framework is plug-and-play and can replace the traditional adversarial domain adaptation module. The proposed method has consistently demonstrated effectiveness through comprehensive experiments on three publicly available datasets, including cross-site prostate segmentation dataset, cross-site COVID-19 lesion segmentation dataset, and cross-modality cardiac segmentation dataset. The source code is available at https://github.com/YonghengSun1997/ODADA.

A Common Neural Account for Social and Nonsocial Decisions.

To date, social and nonsocial decisions have been studied largely in isolation. Consequently, the extent to which social and nonsocial forms of decision uncertainty are integrated using shared neurocomputational resources remains elusive. Here, we address this question using simultaneous electroencephalography (EEG)-functional magnetic resonance imaging (fMRI) in healthy human participants (young adults of both sexes) and a task in which decision evidence in social and nonsocial contexts varies along comparable scales. First, we identify time-resolved build-up of activity in the EEG, akin to a process of evidence accumulation (EA), across both contexts. We then use the endogenous trial-by-trial variability in the slopes of these accumulating signals to construct parametric fMRI predictors. We show that a region of the posterior-medial frontal cortex (pMFC) uniquely explains trial-wise variability in the process of evidence accumulation in both social and nonsocial contexts. We further demonstrate a task-dependent coupling between the pMFC and regions of the human valuation system in dorso-medial and ventro-medial prefrontal cortex across both contexts. Finally, we report domain-specific representations in regions known to encode the early decision evidence for each context. These results are suggestive of a domain-general decision-making architecture, whereupon domain-specific information is likely converted into a "common currency" in medial prefrontal cortex and accumulated for the decision in the pMFC.SIGNIFICANCE STATEMENT Little work has directly compared social-versus-nonsocial decisions to investigate whether they share common neurocomputational origins. Here, using combined electroencephalography (EEG)-functional magnetic resonance imaging (fMRI) and computational modeling, we offer a detailed spatiotemporal account of the neural underpinnings of social and nonsocial decisions. Specifically, we identify a comparable mechanism of temporal evidence integration driving both decisions and localize this integration process in posterior-medial frontal cortex (pMFC). We further demonstrate task-dependent coupling between the pMFC and regions of the human valuation system across both contexts. Finally, we report domain-specific representations in regions encoding the early, domain-specific, decision evidence. These results suggest a domain-general decision-making architecture, whereupon domain-specific information is converted into a common representation in the valuation system and integrated for the decision in the pMFC.

Cross-Lingual Adversarial Domain Adaptation for Novice Programming

Student modeling sits at the epicenter of adaptive learning technology. In contrast to the voluminous work on student modeling for well-defined domains such as algebra, there has been little research on student modeling in programming (SMP) due to data scarcity caused by the unbounded solution spaces of open-ended programming exercises. In this work, we focus on two essential SMP tasks: program classification and early prediction of student success and propose a Cross-Lingual Adversarial Domain Adaptation (CrossLing) framework that can leverage a large programming dataset to learn features that can improve SMP's build using a much smaller dataset in a different programming language. Our framework maintains one globally invariant latent representation across both datasets via an adversarial learning process, as well as allocating domain-specific models for each dataset to extract local latent representations that cannot and should not be united. By separating globally-shared representations from domain-specific representations, our framework outperforms existing state-of-the-art methods for both SMP tasks.

Spatiotemporal and sensory modality attention processing with domain-specific representations in frontoparietal areas.

The frontoparietal network (FPN), including bilateral frontal eye field, inferior parietal sulcus, and supplementary motor area, has been linked to attention processing, including spatiotemporal and sensory modality domains. However, it is unclear whether FPN encodes representations of these domains that are generalizable across subdomains. We decomposed multivariate patterns of functional magnetic resonance imaging activity from 20 participants into domain-specific components and identified latent multivariate representations that generalized across subdomains. The 30 experimental conditions were organized into unimodal-bimodal and spatial-temporal models. We found that brain areas in the FPN, form the primary network that modulated during attention across domains. However, the activation patterns of areas within the FPN were reorganized according to the specific attentional demand, especially when pay attention to different sensory, suggesting distinct regional neural representations associated with specific attentional processes within FPN. In addition, there were also other domain-specific areas outside the FPN, such as the dorsolateral prefrontal cortex. Our conclusion is that, according to the results of the analysis of representation similarity, 2 types of activated brain regions, related to attention domain detailed information processing and general information processing, can be revealed.

Saliency map-guided hierarchical dense feature aggregation framework for breast lesion classification using ultrasound image

Deep learning methods, especially convolutional neural networks, have advanced the breast lesion classification task using breast ultrasound (BUS) images. However, constructing a highly-accurate classification model still remains challenging due to complex pattern, relatively-low contrast and fuzzy boundary existing between lesion regions (i.e., foreground) and the surrounding tissues (i.e., background). Few studies have separated foreground and background for learning domain-specific representations, and then fused them for improving performance of models. In this paper, we propose a saliency map-guided hierarchical dense feature aggregation framework for breast lesion classification using BUS images. Specifically, we first generate saliency maps for foreground and background via super-pixel clustering and multi-scale region grouping. Then, a triple-branch network, including two feature extraction branches and a feature aggregation branch, is constructed to learn and fuse discriminative representations under the guidance of priors provided by saliency maps. In particular, two feature extraction branches take the original image and corresponding saliency map as input for extracting foreground- and background-specific representations. Subsequently, a hierarchical feature aggregation branch receives and fuses the features from different stages of two feature extraction branches, for lesion classification in a task-oriented manner. The proposed model was evaluated on three datasets using 5-fold cross validation, and experimental results have demonstrated that it outperforms several state-of-the-art deep learning methods on breast lesion diagnosis using BUS images.

Similarity of Computations Across Domains Does Not Imply Shared Implementation: The Case of Language Comprehension

Understanding language requires applying cognitive operations (e.g., memory retrieval, prediction, structure building) that are relevant across many cognitive domains to specialized knowledge structures (e.g., a particular language’s lexicon and syntax). Are these computations carried out by domain-general circuits or by circuits that store domain-specific representations? Recent work has characterized the roles in language comprehension of the language network, which is selective for high-level language processing, and the multiple-demand (MD) network, which has been implicated in executive functions and linked to fluid intelligence and thus is a prime candidate for implementing computations that support information processing across domains. The language network responds robustly to diverse aspects of comprehension, but the MD network shows no sensitivity to linguistic variables. We therefore argue that the MD network does not play a core role in language comprehension and that past findings suggesting the contrary are likely due to methodological artifacts. Although future studies may reveal some aspects of language comprehension that require the MD network, evidence to date suggests that those will not be related to core linguistic processes such as lexical access or composition. The finding that the circuits that store linguistic knowledge carry out computations on those representations aligns with general arguments against the separation of memory and computation in the mind and brain.