Low-level Representation Research Articles

Source code plagiarism detection with low-level structural representation and information retrieval

Low-level approach is a recent way for source code plagiarism detection. Instead of relying on source code tokens, it relies on low-level structural representation resulted from compiling given source codes. However, to date, existing low-level approaches are unsuitable for handling large-sized source codes; their comparison takes either quadratic or cubic time complexity. This paper presents a low-level approach with a more-efficient comparison; the comparison only takes linear time complexity with the help of Cosine Correlation in Vector Space Model. According to our evaluation, three findings can be deducted. First, proposed approach can be more efficient and effective than the state-of-the-art approach. Second, low-level structural representation can be more effective and efficient than source code token sequence as a medium for source code plagiarism detection. Third, subsequence length (which is related to n in n-gram) can be inversely proportional to effectiveness without significant impact on efficiency.

Instrumented Tool based Robot Programming - Parameterization of Screwing Process Macros

Programming by Demonstration (PbD), is a powerful mechanism for reducing the complexity of search space for learning. Current approaches for skill representation can be broadly divided between two trends. Low-level representations of skills, taking into account the form of a non-linear mapping between sensory and motor information, which is referred to as “trajectories encoding”. On the other hand high-level representations of skills that decompose them in a sequence of action-perception units which is referred to as “symbolic encoding” [1]. While for symbolic encoding visual perception of states and actions is one preferred approach trajectories usually are acquired from direct demonstration by hand guidance of robot arms carrying robotic tools by skilled workers. This approach requires both the process knowledge of the experienced worker and the possibility of intuitive hand guidance of the robotic tool which is negatively affected by the robots limited working space, friction or the reachability of the robot. Therefore and because it is also highly unintuitive to parameterize macro based robot programs with numerical values considering process parameters (like required forces or position parameters) we propose to let the worker use ordinary power tools during demonstration which are instrumented with additional sensors in order to be able to measure relevant parameters, so called “instrumented tools”. The paper presents the concept for a robotic workstation to be programmed by demonstration with an instrumented tool, the components of the instrumented tool as well as relevant calibration steps and results from application of the technology to an automotive preassembly process..

Cross-Domain Sentiment Classification With Bidirectional Contextualized Transformer Language Models

Cross-domain sentiment classification is an important Natural Language Processing (NLP) task that aims at leveraging knowledge obtained from a source domain to train a high-performance learner for sentiment classification on a target domain. Existing transfer learning methods applied on cross-domain sentiment classification mostly focus on inducing a low-dimensional feature representation shared across domains based on pivots and non-pivots, which is still a low-level representation of sequence data. Recently, there have been great progress in the NLP literature in developing high-level representation language models based on Transformer architecture, which are pre-trained on large text corpus and fine-tuned for specific task with an additional layer on top. Among such language models, the bidirectional contextualized Transformer language models of BERT and XLNet have greatly impacted NLP research field. In this paper, we fine-tune BERT and XLNet for the cross-domain sentiment classification. We then explore their transferability in the context of cross-domain sentiment classification through in-depth analysis of two models' performances and update the state-of-the-arts with a significant margin of improvement. Our results show that such bidirectional contextualized language models outperform the previous state-of-the-arts methods for cross-domain sentiment classification while using up to 120 times less data.

A high-level semantic approach to End-User Development in the Internet of Things

Abstract Various programming environments for End-User Development (EUD) allow the composition of Internet of Things (IoT) applications, i.e., connections between IoT objects to personalize their joint behavior. These environments, however, only support a one-to-one mapping between pairs of object instances, and adopt a low level of abstraction that forces users to be aware of every single technology they may encounter in their applications. As a consequence, numerous open questions remain: would a “higher level” of abstraction help users creating their IoT applications more effectively and efficiently compared with the contemporary low-level representation? Which representation would users prefer? How high-level IoT applications could be actually executed? To answer these questions, we introduce EUPont, a high-level semantic model for EUD in the IoT. EUPont allows the creation of high-level IoT applications, able to adapt to different contextual situations. By integrating the ontology in the architecture of an EUD platform, we demonstrate how the semantic capabilities of the model allow the execution of high-level IoT applications. Furthermore, we evaluate the approach in a user study with 30 participants, by comparing a Web interface for composing IoT applications powered by EUPont with the one employed by a widely used EUD platform. Results show that the high-level approach is understandable, and it allows users to create IoT applications more correctly and quickly than contemporary solutions.

We see the orange not the lemon: typicality effects in ultra-rapid categorization in adults with and without autism spectrum disorder.

Semantic meaning can be extracted from pictures presented very briefly, in the order of tens of milliseconds. This ultra-rapid categorization processing appears to respect a coarse-to-fine path where lower level representations of concepts, or more detailed information, need additional time. We question whether variations in the levels of typicality of the target-item would implicate additional processing for correct classification, both in neurotypical (NT) individuals and with autism spectrum disorder (ASD). Previous research in ASD points out that atypical exemplars of a category might be abnormally processed (e.g., longer times in identifying a penguin as a bird), an observation that we further tested with a rapid serial visual presentation (RSVP) task. In this study, we applied a RSVP task, with four different presentation times (13, 27, 50, and 80ms) and with typical and atypical exemplars to a group of NT individuals and a sample of individuals with ASD. We found, overall, a strong effect of typicality with a higher detection rate for typical items. In addition, we observed a group×typicality×duration interaction. We interpret these findings in the light of the competences of the feedforward sweep of information through our visual system.

Review of Compact Data Structures - a practical approach by Gonzalo Navarro

The book studies fundamental data structures and representations of combinatorial objects, including arrays, bitvectors, permutations, trees, graphs, etc. There are, of course, a great number of existing textbooks and lecture notes about data structures. The perspective of this book is, however, different from the usual presentations. In some sense, this book starts exactly where others leave off saying "we omit the low-level details of representation", and concerns itself mostly with the low-level representation. Specifically, it looks at how to organize and store data structures using as little space as possible while supporting efficient queries and operations. For the considered structures, the book explores the trade-off between reduced space and speed of operation. The analyses of the presented solutions go beyond the usual asymptotics, and also consider exact constant factors (particularly for the leading term of the space complexity); useful practical considerations are also discussed.

Occlusion-Aware Unsupervised Learning of Monocular Depth, Optical Flow and Camera Pose with Geometric Constraints

We present an occlusion-aware unsupervised neural network for jointly learning three low-level vision tasks from monocular videos: depth, optical flow, and camera motion. The system consists of three different predicting sub-networks simultaneously coupled by combined loss terms and is capable of computing each task independently on test samples. Geometric constraints extracted from scene geometry which have traditionally been used in bundle adjustment or pose-graph optimization are formed as various self-supervisory signals during our end-to-end learning approach. Different from prior works, our image reconstruction loss also takes account of optical flow. Moreover, we impose novel 3D flow consistency constraints over the predictions of all the three tasks. By explicitly modeling occlusion and taking utilization of both 2D and 3D geometry relationships, abundant geometric constraints are formed over estimated outputs, enabling the system to capture both low-level representations and high-level cues to infer thinner scene structures. Empirical evaluation on the KITTI dataset demonstrates the effectiveness and improvement of our approach: (1) monocular depth estimation outperforms state-of-the-art unsupervised methods and is comparable to stereo supervised ones; (2) optical flow prediction ranks top among prior works and even beats supervised and traditional ones especially in non-occluded regions; (3) pose estimation outperforms established SLAM systems under comparable input settings with a reasonable margin.

Dissociating parallel and serial processing of numerical value

Digits serve as useful tools for studying the interaction between low-level perceptual representations and higher-level semantic information, and also the degree to which processing these stimulus attributes relies on similar or different mechanisms. Following a body of literature that debates the influence of high-level, semantic information on perceptual processing, and work by Van Opstal and colleagues (2011) investigating whether subliminally presented digit arrays affect estimates of numerical averages, here we explored the temporal dynamics of extracting numerical values associated with each digit in an array. Specifically, we examined reaction times (RTs) for estimating the average of digit arrays of varying sizes to determine whether numerical meaning is extracted in parallel, or instead may require serial processing of individual digits. In Experiment 1, participants completed a numerical mean estimation task, along with visual search tasks designed to yield RT patterns across increasing display sizes thought to be characteristic of serial and parallel processes. In Experiment 2, we controlled for brightness cues that could have contributed to performance in Experiment 1. In both experiments, comparing RT patterns for the numerical averaging tasks to those of the search tasks suggested that semantic information from multiple digits may be extracted by a parallel processing mechanism. Unlike in either search task, RTs improved with increasing display size, indicating a potential benefit of larger displays as has been reported for extracting ensemble, or summary statistical representations of lower-level visual information.

Grid-Based Environment Estimation Using Evidential Mapping and Particle Tracking

Modeling and estimating the current local environment by processing sensor measurement data is essential for intelligent vehicles. Static obstacles, dynamic objects, and free space have to be appropriately represented, classified, and filtered. Occupancy grids, known for mapping static environments, provide a common low-level representation using occupancy probabilities with an implicit data association through the discrete grid structure. Extending this idea toward dynamic environments with moving objects requires a static/dynamic classification of measured occupancy and a tracking of the dynamic state of grid cells. In this paper, we propose a new dynamic grid mapping approach. An evidential representation using the Dempster-Shafer framework is used to model hypotheses for static occupancy, dynamic occupancy, free space, and their combined hypotheses. These hypotheses are consistently estimated and accumulated in a dynamic grid map by an adapted evidential filtering, allowing one to distinguish static and dynamic occupancy. The evidential grid mapping is combined with a low-level particle filter tracking that is used to estimate cell velocity distributions and predict dynamic occupancy of the grid map. Static occupancy is directly modeled in the grid map without requiring particles, increasing efficiency, and improving the static/dynamic classification due to the persistent map accumulation. Experimental results with real sensor data show the effectiveness of the proposed approach in challenging scenarios with occlusions and dense traffic.

Exploiting multiple word embeddings and one-hot character vectors for aspect-based sentiment analysis

Representing words as real-value vectors and use them as input in deep neural networks is an effective approach in many natural language processing tasks. Currently, some studies use a lower-level representation which is character-based vectors. This paper addresses on how to integrate different representations of input for the problem of aspect-based sentiment analysis. We will propose a joint model of multiple Convolutional Neural Networks (CNNs) in which each individual representation of the input is handled by one CNN. In this work we focus on three kinds of representation including word embeddings from the two methods (Word2Vec and GloVe) and the one-hot character vectors. Our experimental results demonstrate that the proposed model can achieve state-of the-art performance in aspect category detection and aspect sentiment classification tasks.

Fluid secretion is significantly reduced in pancreatic ducts from newborn ferrets and pigs with cystic fibrosis

Person re-identification has become a hot research topic due to its importance in surveillance and forensics applications. The purpose is to find the same person from disjoint camera views at different times and locations. Most of the existing approaches try to re-identify a target by modeling his appearance characteristics based on a single modality, i.e. low-level feature or mid-level feature representation. In this paper, we proposed a novel Multi-Level Semantic Appearance Representation that combines these two complementary characteristics for people appearance modeling. The low-level representation relies on the Multi-Channel Co-occurrence Matrix descriptor extracted from the salient body parts to model the fine characteristics of the appearance in terms of color and texture information. The mid-level representation relies on the Semantic Body Traits descriptor to model the semantic description of the appearance in terms of clothes and accessories patterns. A score fusion method based on the weighted sum was adopted to fuse these characteristics that reflect the impact of each separated modality. The experimental results proved the effectiveness and benefits of the proposed approach in terms of performance and computational complexity through the achieved comparisons with the state-of-the-art approaches using the VIPeR benchmarking dataset.

COLORIZING SENTINEL-1 SAR IMAGES USING A VARIATIONAL AUTOENCODER CONDITIONED ON SENTINEL-2 IMAGERY

Abstract. In this paper, we have shown an approach for the automatic colorization of SAR backscatter images, which are usually provided in the form of single-channel gray-scale imagery. Using a deep generative model proposed for the purpose of photograph colorization and a Lab-space-based SAR-optical image fusion formulation, we are able to predict artificial color SAR images, which disclose much more information to the human interpreter than the original SAR data. Future work will aim at further adaption of the employed procedure to our special case of multi-sensor remote sensing imagery. Furthermore, we will investigate if the low-level representations learned intrinsically by the deep network can be used for SAR image interpretation in an end-to-end manner.

Multidimensional brain activity dictated by winner-take-all mechanisms

A novel demon-based architecture is introduced to elucidate brain functions such as pattern recognition during human perception and mental interpretation of visual scenes. Starting from the topological concepts of invariance and persistence, we introduce a Selfridge pandemonium variant of brain activity that takes into account a novel feature, namely, demons that recognize short straight-line segments, curved lines and scene shapes, such as shape interior, density and texture. Low-level representations of objects can be mapped to higher-level views (our mental interpretations): a series of transformations can be gradually applied to a pattern in a visual scene, without affecting its invariant properties. This makes it possible to construct a symbolic multi-dimensional representation of the environment. These representations can be projected continuously to an object that we have seen and continue to see, thanks to the mapping from shapes in our memory to shapes in Euclidean space. Although perceived shapes are 3-dimensional (plus time), the evaluation of shape features (volume, color, contour, closeness, texture, and so on) leads to n-dimensional brain landscapes. Here we discuss the advantages of our parallel, hierarchical model in pattern recognition, computer vision and biological nervous system’s evolution.

Gender-bias in the sensory representation of infant cry

The auditory neural pathway in females appears to be more sensitive to the cry of an infant (De Pisapia et al., 2013; Messina et al., 2016). Cortical responses in females have shown a distinct advantage compared to males in the auditory processing of infant cry. Such gender-bias in the cortical responses might emanate either at higher levels of processing such as cognitive and emotional processing or at the lower level representation of stimulus features. We assessed for a difference if any, between the two genders, in the sensory representation of an infant’s cry. We used frequency following responses (FFR) to assess the sensory representation of an infant cry. This was done in sixteen male and fifteen female non-parent adults. The FFR closely mimics the stimulus acoustics with fine temporal precision and is the measure of choice to assess the sensory encoding of sounds in the auditory system. We performed spectral analysis of the FFRs and compared the spectral magnitudes between males and females. We found significantly higher FFR spectral magnitudes in females compared to males. The gender differences found were not related to the confounding variables such as head size or differences in the volume-conducting media. By systematically controlling other influencing variables, we show that the bias in neural processing of infant cry in females emerges right at the sensory representation levels.

Neural mechanisms of mental schema: a triplet of delta, low beta/spindle and ripple oscillations.

Schemas are higher-level knowledge structures that integrate and organise lower-level representations. As internal templates, schemas are formed according to how events are perceived, interpreted and remembered. Although these higher-level units are assumed to play a fundamental role in our daily life from an early age, the neuronal basis and mechanisms of schema formation and use remain largely unknown. It is important to elucidate how the brain constructs and maintains these higher-level units. In order to examine the possible neural underpinnings of schema, we recapitulate previous work and discuss their findings related to schemas as the brain template. We specifically focused on low beta/spindle oscillations, which are assumed to be the key components of schemas, and propose that the brain template is implemented with a triplet of neural oscillations, that is delta, low beta/spindle and ripple oscillations.

One wouldn't expect an expert bowler to hit only two pins: Hierarchical predictive processing of agent-caused events.

Evidence is accumulating that our brains process incoming information using top-down predictions. If lower level representations are correctly predicted by higher level representations, this enhances processing. However, if they are incorrectly predicted, additional processing is required at higher levels to “explain away” prediction errors. Here, we explored the potential nature of the models generating such predictions. More specifically, we investigated whether a predictive processing model with a hierarchical structure and causal relations between its levels is able to account for the processing of agent-caused events. In Experiment 1, participants watched animated movies of “experienced” and “novice” bowlers. The results are in line with the idea that prediction errors at a lower level of the hierarchy (i.e., the outcome of how many pins fell down) slow down reporting of information at a higher level (i.e., which agent was throwing the ball). Experiments 2 and 3 suggest that this effect is specific to situations in which the predictor is causally related to the outcome. Overall, the study supports the idea that a hierarchical predictive processing model can account for the processing of observed action outcomes and that the predictions involved are specific to cases where action outcomes can be predicted based on causal knowledge.

Towards life-long adaptive agents: using metareasoning for combining knowledge-based planning with situated learning

AbstractWe consider task planning for long-living intelligent agents situated in dynamic environments. Specifically, we address the problem of incomplete knowledge of the world due to the addition of new objects with unknown action models. We propose a multilayered agent architecture that uses meta-reasoning to control hierarchical task planning and situated learning, monitor expectations generated by a plan against world observations, forms goals and rewards for the situated reinforcement learner, and learns the missing planning knowledge relevant to the new objects. We use occupancy grids as a low-level representation for the high-level expectations to capture changes in the physical world due to the additional objects, and provide a similarity method for detecting discrepancies between the expectations and the observations at run-time; the meta-reasoner uses these discrepancies to formulate goals and rewards for the learner, and the learned policies are added to the hierarchical task network plan library for future re-use. We describe our experiments in the Minecraft and Gazebo microworlds to demonstrate the efficacy of the architecture and the technique for learning. We test our approach against an ablated reinforcement learning (RL) version, and our results indicate this form of expectation enhances the learning curve for RL while being more generic than propositional representations.

Evaluating Low-Level Speech Features Against Human Perceptual Data

We introduce a method for measuring the correspondence between low-level speech features and human perception, using a cognitive model of speech perception implemented directly on speech recordings. We evaluate two speaker normalization techniques using this method and find that in both cases, speech features that are normalized across speakers predict human data better than unnormalized speech features, consistent with previous research. Results further reveal differences across normalization methods in how well each predicts human data. This work provides a new framework for evaluating low-level representations of speech on their match to human perception, and lays the groundwork for creating more ecologically valid models of speech perception.

Learning part-based mid-level representation for visual recognition

Abstract There exists a huge semantic gap between the low-level image representations and high-level semantics. To bridge such a gap, this paper proposes a mid-level image representation for visual recognition, where an image is represented based upon the response maps of local part filters. Each dimension of the mid-level representation indicates the likelihood of seeing a part in the input image. The part filters are trained using external data and need not to be fine-tuned on test data. To eliminate the possibly redundant similar parts occurring in different objects or scenes, we perform unsupervised clustering for part refinement. To alleviate the expensive computation of the response maps of the part filters, we further leverage sparse coding to accelerate the feature extraction process, which is ten times faster without significantly compromising the recognition accuracy. We evaluate the proposed mid-level representation on both image and video content recognition tasks and attain state-of-the-art results.

Multi-level semantic appearance representation for person re-identification system

Person re-identification has become a hot research topic due to its importance in surveillance and forensics applications. The purpose is to find the same person from disjoint camera views at different times and locations. Most of the existing approaches try to re-identify a target by modeling his appearance characteristics based on a single modality, i.e. low-level feature or mid-level feature representation. In this paper, we proposed a novel Multi-Level Semantic Appearance Representation that combines these two complementary characteristics for people appearance modeling. The low-level representation relies on the Multi-Channel Co-occurrence Matrix descriptor extracted from the salient body parts to model the fine characteristics of the appearance in terms of color and texture information. The mid-level representation relies on the Semantic Body Traits descriptor to model the semantic description of the appearance in terms of clothes and accessories patterns. A score fusion method based on the weighted sum was adopted to fuse these characteristics that reflect the impact of each separated modality. The experimental results proved the effectiveness and benefits of the proposed approach in terms of performance and computational complexity through the achieved comparisons with the state-of-the-art approaches using the VIPeR benchmarking dataset.