Unstructured Representation Research Articles

CACER: Clinical concept Annotations for Cancer Events and Relations.

Clinical notes contain unstructured representations of patient histories, including the relationships between medical problems and prescription drugs. To investigate the relationship between cancer drugs and their associated symptom burden, we extract structured, semantic representations of medical problem and drug information from the clinical narratives of oncology notes. We present Clinical concept Annotations for Cancer Events and Relations (CACER), a novel corpus with fine-grained annotations for over 48000 medical problems and drug events and 10000 drug-problem and problem-problem relations. Leveraging CACER, we develop and evaluate transformer-based information extraction models such as Bidirectional Encoder Representations from Transformers (BERT), Fine-tuned Language Net Text-To-Text Transfer Transformer (Flan-T5), Large Language Model Meta AI (Llama3), and Generative Pre-trained Transformers-4 (GPT-4) using fine-tuning and in-context learning (ICL). In event extraction, the fine-tuned BERT and Llama3 models achieved the highest performance at 88.2-88.0 F1, which is comparable to the inter-annotator agreement (IAA) of 88.4 F1. In relation extraction, the fine-tuned BERT, Flan-T5, and Llama3 achieved the highest performance at 61.8-65.3 F1. GPT-4 with ICL achieved the worst performance across both tasks. The fine-tuned models significantly outperformed GPT-4 in ICL, highlighting the importance of annotated training data and model optimization. Furthermore, the BERT models performed similarly to Llama3. For our task, large language models offer no performance advantage over the smaller BERT models. We introduce CACER, a novel corpus with fine-grained annotations for medical problems, drugs, and their relationships in clinical narratives of oncology notes. State-of-the-art transformer models achieved performance comparable to IAA for several extraction tasks.

The Emergence of Argument Structure in Artificial Languages

Abstract Computational approaches to the study of language emergence can help us understand how natural languages are shaped by cognitive and sociocultural factors. Previous work focused on tasks where agents refer to a single entity. In contrast, we study how agents predicate, that is, how they express that some relation holds between several entities. We introduce a setup where agents talk about a variable number of entities that can be partially observed by the listener. In the presence of a least-effort pressure, they tend to discuss only entities that are not observed by the listener. Thus we can obtain artificial phrases that denote a single entity, as well as artificial sentences that denote several entities. In natural languages, if we ignore the verb, phrases are usually concatenated, either in a specific order or by adding case markers to form sentences. Our setup allows us to quantify how much this holds in emergent languages using a metric we call concatenability. We also measure transitivity, which quantifies the importance of word order. We demonstrate the usefulness of this new setup and metrics for studying factors that influence argument structure. We compare agents having access to input representations structured into pre-segmented objects with properties, versus unstructured representations. Our results indicate that the awareness of object structure yields a more natural sentence organization.

Effects of Age and Expertise on Mental Representation of the Throwing Movement Among 6- to 16-Year-Olds

The aim of this article was to assess the development of mental representation of the overhead throwing movement as a function of age and expertise. The mental representational structure of the overhead throwing movement was measured using the Structural Dimensional Analysis-Motoric (SDA-M) method that reflects the organization of basic action concepts (BACs). BACs are fundamental building blocks of mental representations, which comprise functional, sensory, spatiotemporal, and biomechanical characteristics of a movement (Schack, 2010). In this study, novices and handball athletes (N = 199) each were grouped according to the level of development in motor ontogenesis (in childhood, pubescence, and adolescents). Male and female handball athletes played in the highest leagues of their age groups. As a result, novices of all age groups showed the same unstructured mental representation. Athletes in the earliest age band resemble all novices’ groups and showed similar unstructured mental representation, whereas athletes within pubescence and adolescents showed functionally well-structured representations, which were similar to the structure of the reference group (N = 8). These results are consistent with a previous investigation of related quantitative and qualitative performance parameters of the overhead throwing movement (Gromeier et al., 2017). Without an increased training, neither the throwing performance nor the associated mental representation is unlikely to improve further by itself or automatically.

Multi-resolution 3D CNN for learning multi-scale spatial features in CAD models

Learning multi-scale spatial features from 3D spatial geometric representations of objects such as point clouds, 3D CAD models, surfaces, and RGB-D data can potentially improve object recognition accuracy. Current deep learning approaches learn such features using structured data representations such as volume occupancy grids (voxels) and octrees or unstructured representations such as graphs and point clouds. Structured representations are generally restricted by their inherent limitations on the resolution, such as the voxel grid dimensions or the maximum octree depth. At the same time, it is challenging to learn directly from unstructured representations of 3D data due to non-uniformity among the samples. A hierarchical approach that maintains the structure at a larger scale while still accounting for the details at a smaller scale in specific spatial locations can provide an optimal solution for learning from 3D data. In this paper, we propose a multi-level learning approach to capture large-scale features at a coarse level (for example, using a coarse voxelization) while simultaneously capturing flexible sparse information of the small-scale features at a fine level (for example, a local fine-level voxel grid) at different spatial locations. To demonstrate the utility of the proposed multi-resolution learning, we use a multi-level voxel representation of CAD models to perform object recognition. The multi-level voxel representation consists of a coarse voxel grid containing volumetric information of the 3D objects and multiple fine-level voxel grids corresponding to each voxel in the coarse grid containing a portion of the object boundary. In addition, we develop an interpretability-based feedback approach to transfer saliency information from one level of features to another in our hierarchical end-to-end learning framework. Finally, we demonstrate the performance of our multi-resolution learning algorithm for object recognition. We outperform several previously published benchmarks for object recognition while using significantly less memory during training.

A Comparative Study of Unstructured Data with SQL and NO-SQL Database Management Systems

This paper aims to establish a relative study between a relational Microsoft SQL Server database and a non-relational MongoDB database within the unstructured representation of data in JSON format. There is a great amount of work done regarding comparison of multiple database management applications on the basis of their performances, security etc., but we have limited information available where these databases are assessed on the basis of provided data. This study will mainly focus on looking at all the possibilities that both these database types offer us when handling data in JSON. We will accomplish this by implementing a series of experiments while taking into consideration that the subjected data does not require to be normalized; and therefore, evaluate the outcome to conclude the result.

Video instance search via spatial fusion of visual words and object proposals

Most popular systems for object instance search are based on the bag-of-visual-word model. The inherent weaknesses of this standard model such as quantization error, unstructured representation, burstiness phenomenon are to some extent solved. However, it has a serious problem of searching small objects on a database with cluttered background. In many situations, even the irrelevant objects which share the same texture or shape with a query object get higher score than relevant ones. To overcome this problem, we propose a novel fusion method to significantly boost the accuracy of instance search systems. Firstly, we use the state-of-the-art object detector with denser feature for finding object bounding box and similarity score. Secondly, to exploit the spatial relationship of each visual word with an object proposal, a detected area that might contain a query object, we define three categories of visual word pairs, i.e., discriminative, weak relevant, and context inferred ones. Finally, we propose a new re-ranking scheme with three weighting functions corresponding to the three categories of visual word pairs to compute the final similarity score between a query topic and a video shot. To illustrate the efficiency of the proposed method, we conduct experiments on datasets which have a wide variety of types of query objects. Experimental results on TRECVID Instance Search datasets (INS2013 and INS2014) show the superiority of our proposed method over the state-of-the-art approaches.

Panoramic Light Field From Hand-Held Video and Its Sampling for Real-Time Rendering

By providing angular and spatial information of light rays, light field images are widely used in many applications. To capture large field of view light field, existing approaches either stitch small field of view light fields, whose apertures are also small, or leverage specialized equipments, which are not accessible to ordinary users. In this paper, we present a method to extract fully 360° field of view panoramic light fields from densely sampled hand-held videos. Unlike previous works, our method handles the large exposure variation and motion blur problems, which are common in the panoramic capture and hand-held videos. To provide real-time rendering performance, our method applies light field sampling to the extracted panoramic light field. Based on the unstructured representation, our method can sample the light field without explicit parameterization. We formulate the sampling problem into a set multicover problem and solve it globally using integer linear programming. Compared with the greedy based method, our approach can use much less video frames to represent the panoramic light field, but still can get better rendering results.

A context aware model for autonomous agent stochastic planning

Markov Decision Processes (MDPs) are not able to make use of domain information effectively due to their representational limitations. The lacking of elements which enable the models be aware of context, leads to unstructured representation of that problem such as raw probability matrices or lists. This causes these tools significantly less efficient at determining a useful policy as the state space of a task grows, which is the case for more realistic problems having localized dependencies between states and actions. In this paper, we present a new state machine, called Context-Aware Markov Decision Process (CA-MDP) based on MDP for the purpose of representing Markovian sequential decision making problems in a more structured manner. CA-MDP changes and augments MDP facilities by integrating causal relationships between actions and states thereby enabling structural, hence compact if possible, representation of the tasks. To show the expressive power of CA-MDP, we give the theoretical bounds for complexity of conversion between MDP and CA-MDP to demonstrate the expressive power of CA-MDP. Next, to generate an optimal policy from CA-MDP encoding by exploiting those newly defined facilities, we devised a new solver algorithm based on value iteration (VI), called Context-Aware Value Iteration (CA-VI). Although regular dynamic programming (DP) based algorithms is successful at effectively determining optimal policies, they do not scale well with respect to state–action space, making both the MDP encoding and related solver mechanism practically unusable for real-life problems. Our solver algorithm gets the power of overcoming the scalability problem by integrating the structural information provided in CA-MDP. First, we give theoretical analysis of CA-VI by examining the expected number of Bellman updates being performed on arbitrary tasks. Finally, we present our conducted experiments on numerous problems, with important remarks and discussions on certain aspects of CA-VI and CA-MDP, to justify our theoretical analyses empirically and to assess the real performance of CA-VI with CA-MDP formulation by analysing the execution time by checking how close it gets to the practical minimum runtime bound with respect to VI performance with MDP encoding of the same task.

A moving mesh unstaggered constrained transport scheme for magnetohydrodynamics

We present a constrained transport (CT) algorithm for solving the 3D ideal magnetohydrodynamic (MHD) equations on a moving mesh, which maintains the divergence-free condition on the magnetic field to machine-precision. Our CT scheme uses an unstructured representation of the magnetic vector potential, making the numerical method simple and computationally efficient. The scheme is implemented in the moving mesh code Arepo. We demonstrate the performance of the approach with simulations of driven MHD turbulence, a magnetized disc galaxy, and a cosmological volume with primordial magnetic field. We compare the outcomes of these experiments to those obtained with a previously implemented Powell divergence-cleaning scheme. While CT and the Powell technique yield similar results in idealized test problems, some differences are seen in situations more representative of astrophysical flows. In the turbulence simulations, the Powell cleaning scheme artificially grows the mean magnetic field, while CT maintains this conserved quantity of ideal MHD. In the disc simulation, CT gives slower magnetic field growth rate and saturates to equipartition between the turbulent kinetic energy and magnetic energy, whereas Powell cleaning produces a dynamically dominant magnetic field. Such difference has been observed in adaptive-mesh refinement codes with CT and smoothed-particle hydrodynamics codes with divergence-cleaning. In the cosmological simulation, both approaches give similar magnetic amplification, but Powell exhibits more cell-level noise. CT methods in general are more accurate than divergence-cleaning techniques, and, when coupled to a moving mesh can exploit the advantages of automatic spatial/temporal adaptivity and reduced advection errors, allowing for improved astrophysical MHD simulations.

A Comparative Study of Databases with Different Methods of Internal Data Management

The purpose of this paper is to present a comparative study between a non-relational MongoDB database and a relational Microsoft SQL Server database in the case of an unstructured representation of data, in XML or JSON format. We mainly focus our presentation on exploring all the possibilities that each type of database offers us, in the case that the data, which has to be stored, cannot or is not wanted to be normalized. This is a scenario most often found in production when, for the application that is being developed we are extracting unstructured data from social networks or all kinds of different channels that the user might have. The comparative study is based on the creation of a benchmark application developed in C# using Visual Studio 2013, which accesses databases created beforehand with proper optimizations that will be described.

Cognitive Architecture and the Epistemic Gap: Defending Physicalism without Phenomenal Concepts

The novel approach presented in this paper accounts for the occurrence of the epistemic gap and defends physicalism against anti-physicalist arguments without relying on so-called phenomenal concepts. Instead of concentrating on conceptual features, the focus is shifted to the special characteristics of experiences themselves. To this extent, the account provided is an alternative to the Phenomenal Concept Strategy. It is argued that certain sensory representations, as accessed by higher cognition, lack constituent structure. Unstructured representations could freely exchange their causal roles within a given system which entails their functional unanalysability. These features together with the encapsulated nature of low level complex processes giving rise to unstructured sensory representations readily explain those peculiarities of phenomenal consciousness which are usually taken to pose a serious problem for contemporary physicalism. I conclude that if those concepts which are related to the phenomenal character of conscious experience are special in any way, their characteristics are derivative of and can be accounted for in terms of the cognitive and representational features introduced in the present paper.

On the representation of task information in task switching: Evidence from task and dimension switching

Task switching research has revealed that task changes lead to a performance switch cost. The present study focuses on the organization of task components in the task set. Three different views of task set organization have been distinguished and evidence in favor of each of these has been reported in the literature. In four experiments, we orthogonally varied the categorization task (magnitude and parity) and the stimulus dimension on which the categorization was to be made. Experiments 1, 2, and 4 used Stroop-like number stimuli, whereas Experiment 3 used global-local stimuli to define the stimulus dimension. In Experiments 2-4, the cue-stimulus interval was also varied. The findings showed that a change of any component resulted in a cost, without any reliable difference in the size of these costs. These results are consistent with the flat view on task-set organization, which assumes that the task set binds all elements in an unstructured representation, which is completely reconfigured each time a change to the task set is required. The implications of these findings are discussed in relation to other findings and the different views on task-set organization.

Transformation of task components into an integrated representation during task switching

We combined task switching with a change paradigm to trace the transformation of the information conveyed by task cues into a readiness to perform the cued task. The tasks were composed of two components that allowed us to determine switch costs as a function of the number of changed task components. With a short stimulus-onset asynchrony (SOA) of 200ms, switch costs grew monotonically as a function of the number of changed components. With a SOA of 400ms, the switch-cost profile exhibited a nonmonotonic characteristic that was previously observed only with switches from an already performed task. Our observations suggest that task preparation can be considered as a process by which an unstructured representation of task components is gradually transformed into an integrated task representation that comprises asymmetrical associations between the task components.

Processus décisionnels de Markov décomposés et factorisés pour l'optimisation de stratégies d'exploration

Autonomous exploration systems require planning under uncertainty. Markov Decision Processes provide a classical framework based on an enumerated and unstructured state space representation. Recent works propose more compact and structured approaches in probabilistic planning. We present and discuss factorization techniques using state variables, and then decomposition techniques using sub-regions. A novel hybrid approach combining both is proposed that is well-fitted to probabilistic exploration-like planning: we use decomposition techniques to generate navigation local policies that are then included in a factored MDP. We discuss results obtained on the basis of probabilistic exploration-like planning problems.

Building model reconstruction based on sensed object location information

The continuous collection of data about the state of facilities appears increasingly feasible due to advances in sensing technologies. In this context, we explore the application of tag-based object location sensing to reconstruct models of existing buildings. We introduce the concept of tag-based building representations. Under certain conditions, these relatively unstructured representations are complete for the automated conversion to boundary-based building representations, which are useful for various construction-related applications. We describe solid modelling routines and spatial queries that facilitate the conversion from tag-based to boundary-based building representations. We provide a case study modelled in a system prototype to demonstrate how the conversion procedure works.

Linear Robust Control of Identified Nonlinear Inverse Compensated SI Engine

A technique based on Horowitz’s nonlinear quantitative feedback theory is presented, in which the size of the uncertainty is reduced by cancellation of the plant nonlinearity through nonlinear inverse compensation. Linear identification of the open loop nonlinear inverse compensated plant is used to obtain an unstructured uncertainty representation, however, rather than computing QFT templates based on parametric uncertainty. An increase in achievable robust performance with any linear robust performance control design technique may then be obtained. The method is experimentally validated by application to a SISO automotive SI engine idle speed control problem using an electric dynamometer.

Linear Robust Control of Identified State-Space Non-Linear Inverse Compensated SI Engine

This paper presents a non-linear inverse compensation methodology based on a technique similar to Horowitz's non-linear quantitative feedback theory. A causal inverse is found based on an identified non-linear plant model. Linear identification of the non-linear identified inverse compensated system is carried out in order to achieve a reduction in the unstructured uncertainty representation. Subsequent robust control design results in an improvement in robust stability and robust performance over the uncompensated system. Using a parameter space linear robust control design technique, the method is applied to the automotive Spark ignition (SI) engine idle speed control problem and is experimentally validated using an electric dynamometer.

Compact Unstructured Representations for Evolutionary Design

This paper proposes a few steps to escape structured extensive representations for objects, in the context of evolutionary Topological Optimum Design (TOD) problems: early results have demonstrated the potential power of Evolutionary methods to find numerical solutions to yet unsolved TOD problems, but those approaches were limited because the complexity of the representation was that of a fixed underlying mesh. Different compact unstructured representations are introduced, the complexity of which is self-adaptive, i.e. is evolved by the algorithm itself. The Voronoi-based representations are variable length lists of alleles that are directly decoded into object shapes, while the IFS representation, based on fractal theory, involves a much more complex morphogenetic process. First results demonstrates that Voronoi-based representations allow one to push further the limits of Evolutionary Topological Optimum Design by actually removing the correlation between the complexity of the representations and that of the discretization. Further comparative results among all these representations on simple test problems seem to indicate that the complex causality in the IFS representation disfavors it compared to the Voronoi-based representations.

Representation and linking mechanisms for audio in MPEG-7

Abstract This paper proposes a general framework for the description of audio within audiovisual sequences for MPEG-7. These related descriptors and description schemes 2 were initially defined during the first phase of MPEG-7 and then evaluated during the Lancaster Meeting held in February 1999. These proposals are based on the underlying premise that audio content can be expressed by a combination of two synergistic representations, both of which are necessary to represent audio content accurately. The first is a structured or semantic representation of audio such as a sentence, paragraph, score, or class. The second is an unstructured representation of the audio simply represented as a continuous stream of data. Since it is not possible to express all aspects of audio in a structured representation, powerful linking mechanisms are required between these two representations. We propose an audio description scheme as a basic structure and representation for audio based on hierarchical, temporal segments. Such a description scheme is essential for both ease of description and to support content based indexing and retrieval of audio. We also propose a description scheme for the representation of larger structures such as spoken content in audio, where the annotation is generated using automatic speech recognition. Finally, we propose linking mechanisms between structured descriptions and unstructured audio content, as an example facility that would add great power to both of the previously mentioned description frameworks.

コネクショニズムと消去主義

Classical computationalism sees cognition as manipulation of syntactically structured representations while connectionism sees it as transformation of syntactically unstructured representations, namely, activation patterns of neurons. J. Fodor and Z. Pylyshyn argue that connectionism fails because every cognitive ability is systematic so that representations in any cognitive domain are syntactically structured. But I argue that some cognitive abilities are not systematic. Classical computationalism holds only for some cognitive domains. But I do not think that our brain is a hybrid of a classical model and a connectionist one. It is wholly connectionist. Syntactically structured representations exist not in our brain but in our environment as external representations. Consequently, eliminativism is right in that propositional attitudes such as belief and desire do not exist in our brain.