Initial Representation Research Articles

Exact learning dynamics of deep linear networks with prior knowledge *

Learning in deep neural networks is known to depend critically on the knowledge embedded in the initial network weights. However, few theoretical results have precisely linked prior knowledge to learning dynamics. Here we derive exact solutions to the dynamics of learning with rich prior knowledge in deep linear networks by generalising Fukumizu’s matrix Riccati solution (Fukumizu 1998 Gen 1 1E–03). We obtain explicit expressions for the evolving network function, hidden representational similarity, and neural tangent kernel over training for a broad class of initialisations and tasks. The expressions reveal a class of task-independent initialisations that radically alter learning dynamics from slow non-linear dynamics to fast exponential trajectories while converging to a global optimum with identical representational similarity, dissociating learning trajectories from the structure of initial internal representations. We characterise how network weights dynamically align with task structure, rigorously justifying why previous solutions successfully described learning from small initial weights without incorporating their fine-scale structure. Finally, we discuss the implications of these findings for continual learning, reversal learning and learning of structured knowledge. Taken together, our results provide a mathematical toolkit for understanding the impact of prior knowledge on deep learning.

Роль рабочей памяти на разных этапах решения инсайтных задач: критический обзор

This paper offers a critical review of studies published over the past few years that explore the role of working memory at different stages of insight problem solving. The overall contribution of working memory (WM) to insight problem solving remains a matter of controversy, since there is supporting evidence for both the positive and the negative roles of executive control and WM in insight problem solving. A promising way to approach this contradiction is to trace WM loading dynamics in the course of a solution. Data analysis revealed that insight problem solving is generally WM-demanding, although to a lesser extent than analytic problem solving. The WM load peaks at the beginning and at the end of insight problem solving. The initial WM load may be linked to the interpretation of a problem description, creating the initial representation, and analytic reasoning within the initial problem space. Subsequently, in the case of insight problems, the WM load decreases noticeably, which can be linked to a search for a new representation or to an impasse. At the last stage of problem solving, a short peak in the WM load precedes the solution detection, which can be linked to the restructuring of representation and the beginning of a new solution process in a new problem space.

Sentimental Contrastive Learning for event representation

Event representation learning is crucial for numerous event-driven tasks, as the quality of event representations greatly influences the performance of these tasks. However, many existing event representation methods exhibit a heavy reliance on semantic features, often neglecting the wealth of information available in other dimensions of events. Consequently, these methods struggle to capture subtle distinctions between events. Incorporating sentimental information can be particularly useful when modeling event data, as leveraging such information can yield superior event representations. To effectively integrate sentimental information, we propose a novel event representation learning framework, namely Sentimental Contrastive Learning (SCL). Specifically, we firstly utilize BERT as the backbone network for pre-training and obtain the initial event representations. Subsequently, we employ instance-level and cluster-level contrastive learning to fine-tune the original event representations. We introduce two distinct contrastive losses respectively for instance-level and cluster-level contrastive learning, each aiming to incorporate sentimental information from different perspectives. To evaluate the effectiveness of our proposed model, we select the event similarity evaluation task and conduct experiments on three representative datasets. Extensive experimental results demonstrate obvious performance improvement achieved by our approach over many other models.

Coarse-to-Fine Knowledge-Enhanced Multi-Interest Learning Framework for Multi-Behavior Recommendation

Multi-types of behaviors (e.g., clicking, carting, purchasing, etc.) widely exist in most real-world recommendation scenarios, which are beneficial to learn users’ multi-faceted preferences. As dependencies are explicitly exhibited by the multiple types of behaviors, effectively modeling complex behavior dependencies is crucial for multi-behavior prediction. The state-of-the-art multi-behavior models learn behavior dependencies indistinguishably with all historical interactions as input. However, different behaviors may reflect different aspects of user preference, which means that some irrelevant interactions may play as noises to the target behavior to be predicted. To address the aforementioned limitations, we introduce multi-interest learning to the multi-behavior recommendation. More specifically, we propose a novel Coarse-to-fine Knowledge-enhanced Multi-interest Learning (CKML) framework to learn shared and behavior-specific interests for different behaviors. CKML introduces two advanced modules, namely Coarse-grained Interest Extracting (CIE) and Fine-grained Behavioral Correlation (FBC) , which work jointly to capture fine-grained behavioral dependencies. CIE uses knowledge-aware information to extract initial representations of each interest. FBC incorporates a dynamic routing scheme to further assign each behavior among interests. Empirical results on three real-world datasets verify the effectiveness and efficiency of our model in exploiting multi-behavior data.

Spontaneous humanization of robots in passive observation of human-robot interaction: A path toward ethical consideration and human-robot cooperation

As humanoid robots become more prevalent, understanding how people form initial representations of and attitudes towards these agents is crucial. This study investigated whether observing a human's emotional reactions towards a robot in a passive human-robot interaction influences the spontaneous humanization and mental state attributions of the robot, and subsequent ethical considerations and cooperation. In Experiment 1, Japanese and British participants observed videos of a human interacting emotionally or neutrally with a Pepper robot and then assessed the robot. Results showed the emotional video elicited more humanization and intentionality attributions towards the robot, subsequently predicting greater ethical concern, especially among Japanese participants. In Experiment 2, observing human emotional engagement increased cooperation with the robot in a Prisoner's Dilemma game through enhanced humanization and intentionality attributions. The findings suggest that passive observation of human-robot interactions can shape the representation, attribution, and treatment of robots based on culture and situational cues. This highlights the need to consider social factors that may influence people's developing relationships with robots.

Query-support semantic correlation mining for few-shot segmentation

Given one or several annotation samples, few-shot segmentation has the capability to complete the segmentation of unseen target objects, thereby reducing the need for extensive data annotation. However, due to the scarcity and diversity of data, it becomes challenging to employ a few support samples effectively for guiding the segmentation of one query sample. Most existing approaches are committed to excavating support samples’ information while ignoring the crucial relationship between query and support samples. To address this limitation and capture richer mutual information, we propose a query-support semantic correlation mining module that serves as an elaborate guide for subsequent dense matching. Specifically, two kinds of query guide maps are generated that leverage global and local similarities to learn the comprehensive correlation between multi-scale support and query features. By incorporating this module, the network can better execute dense matching and mitigate the possibility of misclassification effectively. In addition, a dual-attention module is introduced to obtain a discriminating initial prototype representation to avert prototype bias. Furthermore, our method generalizes well in the k-shot setting. Extensive experiments on PASCAL-5i and COCO-20i datasets demonstrate that the proposed method is potent and successful, which yields competitive segmentation results with state-of-the-art methods.

ARR-GCN: Anatomy-Relation Reasoning Graph Convolutional Network for Automatic Fine-grained Segmentation of Organ's Surgical Anatomy.

Anatomical resection (AR) based on anatomical sub-regions is a promising method of precise surgical resection, which has been proven to improve long-term survival by reducing local recurrence. The fine-grained segmentation of an organ's surgical anatomy (FGS-OSA), i.e., segmenting an organ into multiple anatomic regions, is critical for localizing tumors in AR surgical planning. However, automatically obtaining FGS-OSA results in computer-aided methods faces the challenges of appearance ambiguities among sub-regions (i.e., inter-sub-region appearance ambiguities) caused by similar HU distributions in different sub-regions of an organ's surgical anatomy, invisible boundaries, and similarities between anatomical landmarks and other anatomical information. In this paper, we propose a novel fine-grained segmentation framework termed the "anatomic relation reasoning graph convolutional network" (ARR-GCN), which incorporates prior anatomic relations into the framework learning. In ARR-GCN, a graph is constructed based on the sub-regions to model the class and their relations. Further, to obtain discriminative initial node representations of graph space, a sub-region center module is designed. Most importantly, to explicitly learn the anatomic relations, the prior anatomic-relations among the sub-regions are encoded in the form of an adjacency matrix and embedded into the intermediate node representations to guide framework learning. The ARR-GCN was validated on two FGS-OSA tasks: i) liver segments segmentation, and ii) lung lobes segmentation. Experimental results on both tasks outperformed other state-of-the-art segmentation methods and yielded promising performances by ARR-GCN for suppressing ambiguities among sub-regions.

Spatial-Semantic Collaborative Graph Network for Textbook Question Answering

Textbook Question Answering (TQA) task requires answering questions by reasoning based on both the given diagrams and text context. There are mainly two challenges for the task. First, the diagrams are different from the natural images. Similar shapes or color blocks may express different semantics and there is also a large intra-topic variation for diagrams. Hence, the characteristics of visual semantic ambiguity and variable visual appearance make the diagram understanding more challenging. Second, for the text, the specific education domain with terminologies exists a great gap with the general domain. Therefore, it is difficult to represent the text semantics effectively using a text encoder pretrained in the general domain. In this paper, we propose a Spatial-Semantic Collaborative Graph Network (SSCGN) for TQA task, which can help enhance the diagram and text understanding and facilitate multimodal reasoning. Specifically, the Spatial-guided Semantic Enhancing (SSE) module fully exploits the spatial and semantic relationships between visual objects and OCR tokens to collaboratively enhance the diagram semantic understanding. Moreover, based on the semantically enhanced region representations of the SSE module, the Fine-grained Spatial-Aware Graph Network (FSA-GN) can help obtain richer relation-aware region representations for joint reasoning by capturing more fine-grained spatial relationships. We further propose multiple self-supervised auxiliary tasks to enhance the initial diagram and text semantic representations by pretraining the diagram encoder and text encoder. Extensive experiments and ablation studies are conducted to validate the effectiveness of SSCGN.

20-Month-Old infants’ Use of Noun and Verb Morphosyntactic Cues in Novel Word Learning in Dynamic Events

ABSTRACT Using a habituation paradigm with a three-switch design, the present study investigated whether 20-month-old French-learning infants use noun and verb morphosyntactic cues to learn novel words in dynamic events differentially when both the agent and the action interpretations are possible. Of particular interest was whether infants’ initial interpretation of novel nouns referring to novel animate objects in dynamic events includes not only the novel agents but also their actions. The following two contrastive hypotheses were specifically tested: (1) infants map novel verbs to the novel actions only and novel nouns to the novel agents only. Alternatively, (2) they map novel verbs to the novel actions only but novel nouns to both the novel agents and their actions. Infants watched dynamic events in which novel agents performed novel intransitive actions, while hearing novel words in a noun phrase or a verb sentence. When novel words were preceded by a pronoun “il”, infants were able to map novel verbs to the actions but not to the agents. However, when novel words were preceded by a determiner “un”, they mapped the novel nouns to both the agents and their actions. Two follow-up noun experiments showed that they mapped the novel nouns onto the agents and their actions, even when additional noun morphosyntactic cues were given. These findings demonstrate that 20-month-old infants are able to use noun and verb morphosyntactic cues to learn novel words in dynamic events differentially and provide some evidence to support that infants’ initial representations of novel nouns referring to novel animate objects in dynamic events include both the agents and their actions.

DRGCN: Dynamic Evolving Initial Residual for Deep Graph Convolutional Networks

Graph convolutional networks (GCNs) have been proved to be very practical to handle various graph-related tasks. It has attracted considerable research interest to study deep GCNs, due to their potential superior performance compared with shallow ones. However, simply increasing network depth will, on the contrary, hurt the performance due to the over-smoothing problem. Adding residual connection is proved to be effective for learning deep convolutional neural networks (deep CNNs), it is not trivial when applied to deep GCNs. Recent works proposed an initial residual mechanism that did alleviate the over-smoothing problem in deep GCNs. However, according to our study, their algorithms are quite sensitive to different datasets. In their setting, the personalization (dynamic) and correlation (evolving) of how residual applies are ignored. To this end, we propose a novel model called Dynamic evolving initial Residual Graph Convolutional Network (DRGCN). Firstly, we use a dynamic block for each node to adaptively fetch information from the initial representation. Secondly, we use an evolving block to model the residual evolving pattern between layers. Our experimental results show that our model effectively relieves the problem of over-smoothing in deep GCNs and outperforms the state-of-the-art (SOTA) methods on various benchmark datasets. Moreover, we develop a mini-batch version of DRGCN which can be applied to large-scale data. Coupling with several fair training techniques, our model reaches new SOTA results on the large-scale ogbn-arxiv dataset of Open Graph Benchmark (OGB). Our reproducible code is available on GitHub.

Scalable Attributed-Graph Subspace Clustering

Over recent years, graph convolutional networks emerged as powerful node clustering methods and have set state of the art results for this task. In this paper, we argue that some of these methods are unnecessarily complex and propose a node clustering model that is more scalable while being more effective. The proposed model uses Laplacian smoothing to learn an initial representation of the graph before applying an efficient self-expressive subspace clustering procedure. This is performed via learning a factored coefficient matrix. These factors are then embedded into a new feature space in such a way as to generate a valid affinity matrix (symmetric and non-negative) on which an implicit spectral clustering algorithm is performed. Experiments on several real-world attributed datasets demonstrate the cost-effective nature of our method with respect to the state of the art.

PENDAMPINGAN UMKM ANEKA KRIPIK MELALUI PEMBUATAN LOGO DAN PENGGUNAAN MEDIA ONLINE

One form of Tri Dharma education is community service, which has the purpose of providing insight, skills to become an expertise. The underlying thing to do this pendampngan because local residents are less able to utilize information technology, the problems that exist when sales are less than the target, minimal buyers, the target of this community assistance to the villagers sawahan Puncu and surrounding areas. Logo creation assistance in addition to having a purpose as an initial characteristic when viewing a product through packaging, then make it easier for consumers to provide goods of interest. intended to make it easier for consumers to recognize the products sold. To make an appeal to the public, each logo in a product is considered as an initial representation in recognizing the product. has its own distinctive features. Instagram Facebook, Facebook, tokopedia, shopee, in order to make it easier for manufacturers to sell a product, of course by utilizing appropriate technology, such as using social media instagram, facebook, whatasApp business, marketplace. Marketing using online media makes it easier for MSME actors to reach consumers. Many consumers who appreciate which used to be offline, can now serve orders using online applications. And also payment models are already serving with interbank transfers.

Left atrium surface mesh reconstruction from cardiac MRI

Abstract Funding Acknowledgements Type of funding sources: None. Background The strive towards personalised medicine is universal across all aspects of healthcare. In the management of atrial fibrillation (AF), this can take the form of shape analysis tools to predict an individual’s outcome from catheter ablation (1–3), or electrophysiological simulations to plan an ablation strategy (4–6). Fundamental to their application is our ability to derive accurate three-dimensional (3D) anatomical models of a patient’s left atrial geometry from cross-sectional imaging. MRI is widely utilised in this regard due to its excellent tissue characterisation. However, constructing surface meshes can be challenging given its relatively sparse two-dimensional (2D) image planes. A particular challenge exists in producing a mesh that balances concordance with left atrium segmentation and smoothing terms to eliminate sharp edges and artifact from the contouring process. Objective Here, we describe a novel method for left atrium surface mesh reconstruction from MRI contours. Methods Cardiac MRI was performed on ten patients with paroxysmal or persistent AF, on a 3T clinical MR scanner using phased-array receiver coils. Image sequences included respiratory and ECG-gated 15-minute post contrast late gadolinium enhancement, with a 1.3mm slice thickness and in-plane resolution of 0.66×0.66mm2. Seven individuals were in sinus rhythm, and three in AF at the time of MRI. A single, custom, graphical interface tool implemented in Matlab R2014a enabled 3D mesh reconstruction from 2D MRI slices (7,8). Endocardial surfaces of the left atrium were manually segmented by an experienced clinician, with the pulmonary veins and left atrial appendage excluded at their ostia (Figure 1A). Our tool produced an initial sparse representation with vertices equally spaced along the contours (Figure 1B). Attractor points were defined along the contours, with subdivision and Laplacian smoothing applied to the mesh. This was balanced with a deformation step, which iteratively pulled the mesh towards the attractor points (Figure 1C), resulting in its final shape (Figure 1D). Results Reconstruction performance was evaluated by measuring the distance between heart contours and the reconstructed 3D surface for each case, producing the value for average misalignment 0.71 ± 0.58mm, which compares favourably to the in-plane resolution. Conclusions Our approach offers a highly accurate method for producing 3D anatomical models of the left atrium from MRI, which is fully automated following segmentation through an interactive graphical interface. The iterative optimisation between attractor-based deformation towards the original contours and the Laplacian smoothing enables us to create smooth geometries which remain true to their segmentations. Furthermore, our approach does not require the use of a template mesh, which can result in final geometries being biased towards the template.

Tracing Cognitive Processes in Insight Problem Solving: Using GAMs and Change Point Analysis to Uncover Restructuring

Insight problems are likely to trigger an initial, incorrect mental representation, which needs to be restructured in order to find the solution. Despite the widespread theoretical assumption that this restructuring process happens suddenly, leading to the typical "Aha!" experience, the evidence is inconclusive. Among the reasons for this lack of clarity is that many measures of insight rely solely on the solvers' subjective experience of the solution process. In our previous paper, we used matchstick arithmetic problems to demonstrate that it is possible to objectively trace problem-solving processes by combining eye movements with new analytical and statistical approaches. Specifically, we divided the problem-solving process into ten (relative) temporal phases to better capture possible small changes in problem representation. Here, we go a step further to demonstrate that classical statistical procedures, such as ANOVA, cannot capture sudden representational change processes, which are typical for insight problems. Only nonlinear statistical models, such as generalized additive (mixed) models (GAMs) and change points analysis, correctly identified the abrupt representational change. Additionally, we demonstrate that explicit hints reorient participants' focus in a qualitatively different manner, changing the dynamics of restructuring in insight problem solving. While insight problems may indeed require a sudden restructuring of the initial mental representation, more sophisticated analytical and statistical approaches are necessary to uncover their true nature.

THT: Transfer Head and Tail for Joint Entity Relation Extraction

Extracting entities and relations from unstructured text is the basic task of building knowledge graphs. Various methods based on sequence labeling have been proposed recently, but most of them use uniform token representations to identify entities. It is clear that the first and last words of the head and tail entities should have different representations. For this reason, we propose a method based on word interaction that provides different representations for the first and last words of entities. Specifically, initial token representation is obtained by BERT. Each token interacts with the next k tokens for the entity’s first token recognition, interacts with the previous k tokens for the entity’s last token recognition, and merges them to obtain the head entity. Then recognize the tail entities based on the head entities and form entity pairs. Finally, get the triples based on a simple multi-classification for relation determination. Extensive experimental results on four datasets show that THT performs better than state-of-the-art baselines.

Why do learners overcome non-facilitative transfer faster from an L2 than an L1? The cumulative input threshold hypothesis

ABSTRACT While formal L3 morphosyntactic acquisition research has focused on the roles of the L1 versus L2 during the L3 initial stages, we examine their roles during development. Specifically, we explore whether the L3 (here, Portuguese) revision process after non-facilitative transfer differs whether the initial L3 representation reflects the L1 or L2. If this initial L3 representation is the same for a group whose source is their L1 and a group whose source is their L2, it is plausible to predict similar developmental processes between groups. However, recent data have shown between-group developmental differences despite a similar L3 starting point. In an effort to isolate the variable(s) behind such differences, we investigate the roles of dominance and formal instruction via two sets of L3 morphosyntactic judgments. The first set excludes dominance, with similar patterns in English-dominant versus Spanish-dominant L1 Spanish data. The second set excludes formal instruction, whereby L2 Spanish data shows L3 convergence for a property not explicitly taught and L1 Spanish data does not. As an initial hypothesis for further exploration of between-group differences, we propose the Cumulative Input Threshold Hypothesis: The rate of developmental recovery from non-facilitative transfer is inversely related to lifetime input of the source language.

Multi-hypothesis representation learning for transformer-based 3D human pose estimation

Despite significant progress, estimating 3D human poses from monocular videos remains a challenging task due to depth ambiguity and self-occlusion. Most existing works attempt to solve both issues by exploiting spatial and temporal relationships. However, those works ignore the fact that it is an inverse problem where multiple feasible solutions (i.e., hypotheses) exist. To relieve this limitation, we propose a Multi-Hypothesis Transformer that learns spatio-temporal representations of multiple plausible pose hypotheses. In order to effectively model multi-hypothesis dependencies and build strong relationships across hypothesis features, we introduce a one-to-many-to-one three-stage framework: (i) Generate multiple initial hypothesis representations; (ii) Model self-hypothesis communication, merge multiple hypotheses into a single converged representation and then partition it into several diverged hypotheses; (iii) Learn cross-hypothesis communication and aggregate the multi-hypothesis features to synthesize the final 3D pose. Through the above processes, the final representation is enhanced and the synthesized pose is much more accurate. Extensive experiments show that the proposed method achieves state-of-the-art results on two challenging datasets: Human3.6M and MPI-INF-3DHP. The code and models are available at https://github.com/Vegetebird/MHFormer.

Co-Visual Pattern-Augmented Generative Transformer Learning for Automobile Geo-Localization

Geolocation is a fundamental component of route planning and navigation for unmanned vehicles, but GNSS-based geolocation fails under denial-of-service conditions. Cross-view geo-localization (CVGL), which aims to estimate the geographic location of the ground-level camera by matching against enormous geo-tagged aerial (e.g., satellite) images, has received a lot of attention but remains extremely challenging due to the drastic appearance differences across aerial–ground views. In existing methods, global representations of different views are extracted primarily using Siamese-like architectures, but their interactive benefits are seldom taken into account. In this paper, we present a novel approach using cross-view knowledge generative techniques in combination with transformers, namely mutual generative transformer learning (MGTL), for CVGL. Specifically, by taking the initial representations produced by the backbone network, MGTL develops two separate generative sub-modules—one for aerial-aware knowledge generation from ground-view semantics and vice versa—and fully exploits the entirely mutual benefits through the attention mechanism. Moreover, to better capture the co-visual relationships between aerial and ground views, we introduce a cascaded attention masking algorithm to further boost accuracy. Extensive experiments on challenging public benchmarks, i.e., CVACT and CVUSA, demonstrate the effectiveness of the proposed method, which sets new records compared with the existing state-of-the-art models. Our code will be available upon acceptance.

Обеспечение доступности текстовых материалов в процессе обучения детей с РАС в начальной школе

<p style="text-align: justify;">Children with autism spectrum disorder (ASD) are often limited in understanding spoken and written language. To make texts accessible for children with ASD specialists use means that improve understanding of verbal or written information and make learning process more successful. The article discusses why children with ASD have difficulties in precepting and understanding textual information, for example, poor vocabulary, ignorance of words lexical meaning, misunderstanding of figurative language or hidden meaning behind words due to poor social experience. We present the main working areas to develop and improve speech understanding: improving reading skills, developing the ability to listen to an interlocutor, developing phrasal speech, forming initial grammatical representations, etc. The main text types that children encounter in the classroom are also listed, for example, instructions, rules, tasks, literary works. We also offer tools that facilitate working with texts, make them more accessible, and adapt to individual capabilities of students. We give practical recommendations that will be useful both to teachers working with students with ASD and their parents.</p>

The Use of Attentive Knowledge Graph Perceptual Propagation for Improving Recommendations

Collaborative filtering (CF) usually suffers from data sparsity and cold starts. Knowledge graphs (KGs) are widely used to improve recommendation performance. To verify that knowledge graphs can further alleviate the above problems, this paper proposes an end-to-end framework that uses attentive knowledge graph perceptual propagation for recommendations (AKGP). This framework uses a knowledge graph as a source of auxiliary information to extract user–item interaction information and build a sub-knowledge base. The fusion of structural and contextual information is used to construct fine-grained knowledge graphs via knowledge graph embedding methods and to generate initial embedding representations. Through multi-layer propagation, the structured information and historical preference information are embedded into a unified vector space, and the potential user–item vector representation is expanded. This article used a knowledge perception attention module to achieve feature representation, and finally, the model was optimized using the stratified sampling joint learning method. Compared with the baseline model using MovieLens-1M, Last-FM, Book-Crossing and other data sets, the experimental results demonstrate that the model outperforms state-of-the-art KG-based recommendation methods, and the shortcomings of the existing model are improved. The model was applied to product design data and historical maintenance records provided by an automotive parts manufacturing company. The predictions of the recommended system are matched to the product requirements and possible failure records. This helped reduce costs and increase productivity, helping the company to quickly determine the cause of failures and reduce unplanned downtime.