General Prior Knowledge Research Articles

To Boost Zero-Shot Generalization for Embodied Reasoning With Vision-Language Pre-Training.

Recently, there exists an increased research interest in embodied artificial intelligence (EAI), which involves an agent learning to perform a specific task when dynamically interacting with the surrounding 3D environment. There into, a new challenge is that many unseen objects may appear due to the increased number of object categories in 3D scenes. It makes developing models with strong zero-shot generalization ability to new objects necessary. Existing work tries to achieve this goal by providing embodied agents with massive high-quality human annotations closely related to the task to be learned, while it is too costly in practice. Inspired by recent advances in pre-trained models in 2D visual tasks, we attempt to boost zero-shot generalization for embodied reasoning with vision-language pre-training that can encode common sense as general prior knowledge. To further improve its performance on a specific task, we rectify the pre-trained representation through masked scene graph modeling (MSGM) in a self-supervised manner, where the task-specific knowledge is learned from iterative message passing. Our method can improve a variety of representative embodied reasoning tasks by a large margin (e.g., over 5.0% w.r.t. answer accuracy on MP3D-EQA dataset that consists of many real-world scenes with a large number of new objects during testing), and achieve the new state-of-the-art performance.

Customized design of amorphous solids by generative deep learning

<p>The design of advanced amorphous solids, such as metallic glasses, with targeted properties through artificial intelligence signifies a paradigmatic shift in physical metallurgy and materials technology. Here, we developed a machine learning architecture that facilitates the generation of metallic glasses with targeted multifunctional properties. Our architecture integrates the state-of-the-art unsupervised generative adversarial network model with supervised models, allowing the incorporation of general prior knowledge, derived from thousands of data points across a vast range of alloy compositions, into the creation of data points for a specific type of composition, which overcame the common issue of data scarcity typically encountered in the design of a given type of metallic glasses. Using our generative model, we have successfully designed copper-based metallic glasses, which display exceptionally high hardness or a remarkably low modulus. Notably, our architecture can not only explore uncharted regions in the targeted compositional space but also permits self-improvement after experimental validated data points are added to the initial dataset for subsequent cycles of data generation, hence paving the way for the customized design of amorphous solids without human intervention.</p>

The Motivational Utility of Knowledge: Examining Fundamental Needs in the Context of Houselessness Knowledge

Past research on knowledge has differentiated between dimensions (e.g., amount, accuracy, specificity, coherence) of knowledge. This paper introduces a novel dimension of knowledge, the Motivational Utility of Knowledge (MUK), that is based on fundamental human needs (e.g., physical safety, affiliation, actualization, reproduction). Adults in the United States (N = 190) were recruited from an online survey platform and paid for participation. Participants read a set of four texts arguing different views of houselessness and were administered a comprehension test after each text. Participants were asked about their conceptions of houselessness before and after reading. Finally, they were given the MUK scale, a demographics questionnaire, including questions about their personal experience with houselessness, and were administered a general prior knowledge test and a vocabulary knowledge test. We examined MUK, the factor structure of the scale and the relationship between MUK and other measures of knowledge. The analyses showed that the subscales of MUK loaded onto a single factor—an overall value of houselessness knowledge. In addition, we found that MUK was correlated with conceptions of houselessness and comprehension of texts on houselessness, indicating that the scale was valid. Overall, the findings demonstrate that MUK is an important dimension of knowledge to consider in learning tasks.

Model‐based Crowd Behaviours in Human‐solution Space

AbstractRealistic crowd simulation has been pursued for decades, but it still necessitates tedious human labour and a lot of trial and error. The majority of currently used crowd modelling is either empirical (model‐based) or data‐driven (model‐free). Model‐based methods cannot fit observed data precisely, whereas model‐free methods are limited by the availability/quality of data and are uninterpretable. In this paper, we aim at taking advantage of both model‐based and data‐driven approaches. In order to accomplish this, we propose a new simulation framework built on a physics‐based model that is designed to be data‐friendly. Both the general prior knowledge about crowds encoded by the physics‐based model and the specific real‐world crowd data at hand jointly influence the system dynamics. With a multi‐granularity physics‐based model, the framework combines microscopic and macroscopic motion control. Each simulation step is formulated as an energy optimization problem, where the minimizer is the desired crowd behaviour. In contrast to traditional optimization‐based methods which seek the theoretical minimizer, we designed an acceleration‐aware data‐driven scheme to compute the minimizer from real‐world data in order to achieve higher realism by parameterizing both velocity and acceleration. Experiments demonstrate that our method can produce crowd animations that are more realistically behaved in a variety of scales and scenarios when compared to the earlier methods.

Prediction of Compound Plasma Concentration-Time Profiles in Mice Using Random Forest.

Pharmacokinetic (PK) parameters such as clearance (CL) and volume of distribution (Vd) have been the subject of previous in silico predictive models. However, having information of the concentration over time profile explicitly can provide additional value like time above MIC or AUC, etc., to understand both the efficacy and safety-related aspects of a compound. In this work, we developed machine learning models for plasma concentration-time profiles after both i.v. and p.o. dosing for a series of 17 in-house projects. For explanatory variables, MACCS Keys chemical descriptors as well as in silico and experimental in vitro PK parameters were used. The predictive accuracy of random forest (RF), message passing neural network, 2-compartment models using estimated CL and Vdss, and an average model (as a control experiment) was investigated using 5-fold cross-validation (5-fold CV) and leave-one-project-out validation (LOPO-V). The predictive accuracy of RF in 5-fold CV for i.v. and p.o. plasma concentration-time profiles was the best among the models studied, with an RMSE for i.v. dosing at 0.08, 1, and 8 h of 0.245, 0.474, and 0.462, respectively, and an RMSE for p.o. dosing at 0.25, 1, and 8 h of 0.500, 0.612, and 0.509, respectively. Furthermore, by investigating the importance of the in vitro PK parameters using the Gini index, we observed that the general prior knowledge in ADME research was reflected well in the respective feature importance of in vitro parameters such as predicted human Vd (hVd) for the initial distribution, mouse intrinsic CL and unbound fraction of mouse plasma for the elimination process, and Caco2 permeability for the absorption process. Also, this model is the first model that can predict twin peaks in the concentration-time profile much better than a baseline compartment model. Because of its combination of sufficient accuracy and speed of prediction, we found the model to be fit-for-purpose for practical lead optimization.

Recollection and prior knowledge recruit the left angular gyrus during recognition.

The human angular gyrus (AG) is implicated in recollection, or the ability to retrieve detailed memory content from a specific episode. A separate line of research examining the neural bases of more general mnemonic representations, extracted over multiple episodes, also highlights the AG as a core region of interest. To reconcile these separate views of AG function, the present fMRI experiment used a Remember-Know paradigm with famous (prior knowledge) and non-famous (no prior knowledge) faces to test whether AG activity could be modulated by both task-specific recollection and general prior knowledge within the same individuals. Increased BOLD activity in the left AG was observed during both recollection in the absence of prior knowledge (recollected > non-recollected or correctly rejected non-famous faces) and when prior knowledge was accessed in the absence of experiment-specific recollection (famous > non-famous correct rejections). This pattern was most prominent for the left AG as compared to the broader inferior parietal lobe. Recollection-related responses in the left AG increased with encoding duration and prior knowledge, despite prior knowledge being incidental to the recognition decision. Overall, the left AG appears sensitive to both task-specific recollection and the incidental access of general prior knowledge, thus broadening our notions of the kinds of mnemonic representations that drive activity in this region.

Knowledge-driven feature component interpretable network for motor imagery classification

Objective. The end-to-end convolutional neural network (CNN) has achieved great success in motor imagery (MI) classification without a manual feature design. However, all the existing deep network solutions are purely datadriven and lack interpretability, which makes it impossible to discover insightful knowledge from the learned features, not to mention to design specific network structures. The heavy computational cost of CNN also makes it challenging for real-time application along with high classification performance. Approach. To address these problems, a novel knowledge-driven feature component interpretable network (KFCNet) is proposed, which combines spatial and temporal convolution in analogy to independent component analysis and a power spectrum pipeline. Prior frequency band knowledge of sensory-motor rhythms has been formulated as band-pass linear-phase digital finite impulse response filters to initialize the temporal convolution kernels to enable the knowledge-driven mechanism. To avoid signal distortion and achieve a linear phase and unimodality of filters, a symmetry loss is proposed, which is used in combination with the cross-entropy classification loss for training. Besides the general prior knowledge, subject-specific time-frequency property of event-related desynchronization and synchronization has been employed to construct and initialize the network with significantly fewer parameters. Main results. Comparison of experiments on two public datasets has been performed. Interpretable feature components could be observed in the trained model. The physically meaningful observation could efficiently assist the design of the network structure. Excellent classification performance on MI has been obtained. Significance. The performance of KFCNet is comparable to the state-of-the-art methods but with much fewer parameters and makes real-time applications possible.

Analyzing Firm Reports for Volatility Prediction: A Knowledge-Driven Text-Embedding Approach

Predicting stock return volatility is the key to investment and risk management. Traditional volatility-forecasting methods primarily rely on stochastic models. More recently, many machine-learning approaches, particularly text-mining techniques, have been implemented to predict stock return volatility, thus taking advantage of the availability of large amounts of unstructured data such as firm financial reports. Most existing studies develop simple but effective models to analyze text, such as dictionary-based matching algorithms that use a set of manually constructed keywords. However, the latent and deep semantics encoded in text are usually neglected. In this study, we build on recent progress in representation learning and propose a novel word-embedding method that incorporates external knowledge from a well-known finance-domain lexicon (the Loughran and McDonald (2011) word list), which helps us learn semantic relationships among words in firm reports for better volatility prediction. Using over 10 years of annual reports from Russell 3000 firms, we empirically show that, compared with cutting-edge benchmarks, our proposed method achieves significant improvement in terms of prediction error, for example, a 28.4% reduction on average. We also discuss the practical and methodological implications of our findings. Our financial-specific word-embedding program is available as open-source information so that researchers can use it to analyze financial reports and assess financial risks. Summary of Contribution: Predicting stock return volatility is the key to investment and risk management. Traditional volatility-forecasting methods primarily rely on stochastic models. More recently, many machine-learning, especially text-mining, techniques have been developed to predict stock return volatility given the availability of a large amount of unstructured data, such as firm annual reports. Most existing research develops simple but effective approaches, for example, manually constructing a set of keywords to analyze texts. However, the latent and deep semantics encoded in texts are usually ignored. In this research, we build on recent progress in representation learning and propose a novel word-embedding method that incorporates external knowledge from the finance-domain lexicon of Loughran and McDonald (2011), which helps us learn the semantic relationships among words in firm annual reports for better volatility prediction. In this study, we make the following contributions. First, methodologically, we are among the first to incorporate finance-specific lexicon into representation learning for stock volatility prediction. We propose a novel knowledge-driven text-embedding model that is trained on a large amount of unstructured textual data to learn high quality word embedding. Our proposed approach is effective in predicting stock return volatility, and the approach can potentially have broader applications. Second, substantively, we empirically show that the domain lexicon enhanced text representation learning can indeed significantly improve the performance, compared with bag-of-words models and generic word embedding for volatility prediction. Domain knowledge combined with text learning plays a critical enabling role in understanding financial reports. Third, our method adds on to existing literature on designing financial information systems by incorporating ontology knowledge, common-sense knowledge, and general prior knowledge.

Sparsity-Based Image Inpainting Detection via Canonical Correlation Analysis With Low-Rank Constraints

Image inpainting, a commonly used image editing technique for filling the mask or missing areas in images, is often adopted to destroy the integrity of images by forgers with ulterior motives. Compared with the other types of inpainting, the sparsity-based inpainting exploits more general prior knowledge and has a broader application scope. Although many methods for detecting exemplar-based and diffusion-based inpainting have been successfully studied in the literature, there is still a lack of effective schemes for detecting the sparsity-based inpainting. In this paper, to fill this gap, we proposed a novel algorithm for sparsity-based image inpainting detection. We revealed the potential connection between sparsity-based inpainting and canonical correlation analysis (CCA). This type of inpainting has a strong effect on the CCA coefficients. Based on this observation, a modified objective function of CCA and a corresponding optimization algorithm are further proposed to enhance the inter-class difference in our feature set. Experimental results on three publicly available data sets demonstrated our method’s superiority over other competitors. Particularly, compared with previous inpainting detection methods, the proposed framework yields better performances in the cases of JPEG compression and Gaussian noise addition. The proposed method also shows promising results when employed to detect other types of inpainting.

Abstract 3883: Variant peptide identification system for bottom-up proteomics: Finding hidden sequence alterations in MS data

Abstract In database search, acquired spectra are matched with theoretical spectra of database peptides. Peptides from large databases are likely to match to more than one spectrum, but only a small proportion of peptide interpretations are correct. Incorrect interpretations are likely to arise for two reasons: random matches and non-random matches to homologous peptides. The former is often eliminated by selection of an appropriate similarity threshold or by utilization of the statistical significance of a spectral match. The match to homologous peptide poses a challenge as there is often large similarity in their fragmentation spectra as well. Non-random, but incorrect matches occur surprisingly often in the identification of peptide sequence variants. For this reason, a high identification score does not imply correctness of interpretation. Increasing spectral score threshold enhances specificity in general, but results in significant loss of sensitivity. Therefore, a single spectral criterion is inadequate for distinguishing between homologous peptides. In this study, we reconceive the system for elimination of incorrect interpretations. We propose a system that uses multidimensional LC/MS information in combination with a priori knowledge of sample content to filter out unlikely interpretations. The system calculates retention time of candidate peptides, performs isotopic analysis of precursors and constructs preliminary protein assembly to increase separation between correct and incorrect matches. The approach is used for reevaluation of database search results and is generally applicable to the analysis of standard bottom-up proteomic data. The system was used for identification of alterations in MS2 proteomics data and validated against RNA-Seq data from the same sample. The results demonstrate that non-spectral data can be used to efficiently eliminate peptide interpretations that have no correspondence in the RNA and as such are likely false positives. The approach is sensitive and yields a large proportion of altered peptides that have RNA-Seq support. The method can potentially help overcome the problem of the large database search present, e.g., in proteogenomic studies, particularly important in cancer research and diagnostics. In summary, general prior knowledge of sample content and the use of LC and MS1 data improve on the MS2-based identification of peptides. Citation Format: Marian Hajduch, Miroslav Hruska, Lakshman Varanasi, Jiri Voller, Petr Dzubak. Variant peptide identification system for bottom-up proteomics: Finding hidden sequence alterations in MS data. [abstract]. In: Proceedings of the 107th Annual Meeting of the American Association for Cancer Research; 2016 Apr 16-20; New Orleans, LA. Philadelphia (PA): AACR; Cancer Res 2016;76(14 Suppl):Abstract nr 3883.

Weakly Supervised Large Scale Object Localization with Multiple Instance Learning and Bag Splitting.

Localizing objects of interest in images when provided with only image-level labels is a challenging visual recognition task. Previous efforts have required carefully designed features and have difficulty in handling images with cluttered backgrounds. Up-scaling to large datasets also poses a challenge to applying these methods to real applications. In this paper, we propose an efficient and effective learning framework called MILinear, which is able to learn an object localization model from large-scale data without using bounding box annotations. We integrate rich general prior knowledge into a learning model using a large pre-trained convolutional network. Moreover, to reduce ambiguity in positive images, we present a bag-splitting algorithm that iteratively generates new negative bags from positive ones. We evaluate the proposed approach on the challenging Pascal VOC 2007 dataset, and our method outperforms other state-of-the-art methods by a large margin; some results are even comparable to fully supervised models trained with bounding box annotations. To further demonstrate scalability, we also present detection results on the ILSVRC 2013 detection dataset, and our method outperforms supervised deformable part-based model without using box annotations.

Semantic-based regularization for learning and inference

This paper proposes a unified approach to learning from constraints, which integrates the ability of classical machine learning techniques to learn from continuous feature-based representations with the ability of reasoning using higher-level semantic knowledge typical of Statistical Relational Learning. Learning tasks are modeled in the general framework of multi-objective optimization, where a set of constraints must be satisfied in addition to the traditional smoothness regularization term. The constraints translate First Order Logic formulas, which can express learning-from-example supervisions and general prior knowledge about the environment by using fuzzy logic. By enforcing the constraints also on the test set, this paper presents a natural extension of the framework to perform collective classification. Interestingly, the theory holds for both the case of data represented by feature vectors and the case of data simply expressed by pattern identifiers, thus extending classic kernel machines and graph regularization, respectively. This paper also proposes a probabilistic interpretation of the proposed learning scheme, and highlights intriguing connections with probabilistic approaches like Markov Logic Networks. Experimental results on classic benchmarks provide clear evidence of the remarkable improvements that are obtained with respect to related approaches.

Scene transformation for detector adaptation

This paper focuses on detecting vehicles in different target scenes with the same pre-trained detector which is very challenging due to view variations. To address this problem, we propose a novel approach for detection adaptation based on scene transformation, which contributes in both view transformation and automatic parameter estimation. Instead of modifying the pre-trained detectors, we transform scenes into frontal/rear view handling with pitch and yaw view variations. Without human interactions but only some general prior knowledge, the transformation parameters are automatically initialized, and then online optimized with spatial–temporal voting, which guarantees that the transformation matches the pre-trained detector. Since there is no need of labeling new samples and manual camera calibration, our approach can considerably reduce manual interactions. Experiments on challenging real-world videos demonstrate that our approach achieves significant improvements over the pre-trained detector, and it is even comparable to the performance of the detector trained on fully labeled sequences.

Automated Lipid A Structure Assignment from Hierarchical Tandem Mass Spectrometry Data

Infusion-based electrospray ionization (ESI) coupled to multiple-stage tandem mass spectrometry (MS(n)) is a standard methodology for investigating lipid A structural diversity (Shaffer et al. J. Am. Soc. Mass. Spectrom. 18(6), 1080-1092, 2007). Annotation of these MS(n) spectra, however, has remained a manual, expert-driven process. In order to keep up with the data acquisition rates of modern instruments, we devised a computational method to annotate lipid A MS(n) spectra rapidly and automatically, which we refer to as hierarchical tandem mass spectrometry (HiTMS) algorithm. As a first-pass tool, HiTMS aids expert interpretation of lipid A MS(n ) data by providing the analyst with a set of candidate structures that may then be confirmed or rejected. HiTMS deciphers the signature ions (e.g., A-, Y-, and Z-type ions) and neutral losses of MS(n) spectra using a species-specific library based on general prior structural knowledge of the given lipid A species under investigation. Candidates are selected by calculating the correlation between theoretical and acquired MS(n) spectra. At a false discovery rate of less than 0.01, HiTMS correctly assigned 85% of the structures in a library of 133 manually annotated Francisella tularensis subspecies novicida lipid A structures. Additionally, HiTMS correctly assigned 85% of the structures in a smaller library of lipid A species from Yersinia pestis demonstrating that it may be used across species.

Training signaling pathway maps to biochemical data with constrained fuzzy logic: quantitative analysis of liver cell responses to inflammatory stimuli.

Predictive understanding of cell signaling network operation based on general prior knowledge but consistent with empirical data in a specific environmental context is a current challenge in computational biology. Recent work has demonstrated that Boolean logic can be used to create context-specific network models by training proteomic pathway maps to dedicated biochemical data; however, the Boolean formalism is restricted to characterizing protein species as either fully active or inactive. To advance beyond this limitation, we propose a novel form of fuzzy logic sufficiently flexible to model quantitative data but also sufficiently simple to efficiently construct models by training pathway maps on dedicated experimental measurements. Our new approach, termed constrained fuzzy logic (cFL), converts a prior knowledge network (obtained from literature or interactome databases) into a computable model that describes graded values of protein activation across multiple pathways. We train a cFL-converted network to experimental data describing hepatocytic protein activation by inflammatory cytokines and demonstrate the application of the resultant trained models for three important purposes: (a) generating experimentally testable biological hypotheses concerning pathway crosstalk, (b) establishing capability for quantitative prediction of protein activity, and (c) prediction and understanding of the cytokine release phenotypic response. Our methodology systematically and quantitatively trains a protein pathway map summarizing curated literature to context-specific biochemical data. This process generates a computable model yielding successful prediction of new test data and offering biological insight into complex datasets that are difficult to fully analyze by intuition alone.

Reproducibility of cerebral phenylalanine levels in patients with phenylketonuria determined by 1H‐MR spectroscopy

The reproducibility of metabolite content determined by MR spectroscopy (MRS) is usually at best a few percent for the prominent singlets. When studying low-concentration metabolites, like phenylalanine (Phe), where tissue content can be <100 micromol/kg, better reproducibility is paramount-particularly in view of using MRS results for potential individual treatment advice. An optimized, targeted spectroscopy method was established at 1.5T and reproducibility was established in 21 patients with phenylketonuria (PKU) where three spectra were recorded in each of three independent sessions, two of which were in immediate succession to minimize physiologic variation. Intersession variation was found to be only 7 micromol/kg Phe for back-to-back repetition of sessions, in close agreement with the variation of 16 micromol/kg observed for single spectra within a session. Analysis of variance proved the individuality of the blood/brain Phe ratio-though this ratio seems to be influenced by physiologic factors that are not stable in time. The excellent reproducibility was achieved through optimization of various factors, including signal-to-noise ratio, repositioning, and prescan calibrations, but also by enforcing as much prior information as possible (e.g., lineshape and phase from reference scans, constant prior-knowledge-locked baseline). While the application of maximum general prior knowledge is a general method to reduce fluctuations, one should remember that it may introduce systematic errors.

EM enhancement of 3D head pose estimated by point at infinity

Head pose estimation is a key task for visual surveillance, HCI and face recognition applications. In this paper, a new approach is proposed for estimating 3D head pose from a monocular image. The approach assumes the full perspective projection camera model. Our approach employs general prior knowledge of face structure and the corresponding geometrical constraints provided by the location of a certain vanishing point to determine the pose of human faces. To achieve this, eye-lines, formed from the far and near eye corners, and mouth-line of the mouth corners are assumed parallel in 3D space. Then the vanishing point of these parallel lines found by the intersection of the eye-line and mouth-line in the image can be used to infer the 3D orientation and location of the human face. In order to deal with the variance of the facial model parameters, e.g. ratio between the eye-line and the mouth-line, an EM framework is applied to update the parameters. We first compute the 3D pose using some initially learnt parameters (such as ratio and length) and then adapt the parameters statistically for individual persons and their facial expressions by minimizing the residual errors between the projection of the model features points and the actual features on the image. In doing so, we assume every facial feature point can be associated to each of features points in 3D model with some a posteriori probability. The expectation step of the EM algorithm provides an iterative framework for computing the a posterori probabilities using Gaussian mixtures defined over the parameters. The robustness analysis of the algorithm on synthetic data and some real images with known ground-truth are included.

Pose determination of human faces by using vanishing points

A new method for estimating 3D-head pose from a monocular image is proposed in this paper. The approach employs general prior knowledge of face structure and the corresponding geometrical constraints provided by the location of vanishing point to determine pose of human faces. The connection of the two far-eye corners and the two neighboring-eye corners, respectively, form the eye-lines. Connecting the two far-mouth corners forms the mouth-line. The eye-lines and the mouth-line are assumed parallel in 3D space, and their vanishing point on the image plane can be used to infer 3D pose of the human face. Perspective projection imaging model is used and an accurate analytic solution of pose (position and orientation) of human face is deduced. The orientation of the facial plane can be obtained when the ratio of the lengths of the eye-line segment (far-eye corners) and the mouth-line segment (far-mouth corners) is known. Furthermore, if one of the two lengths is known, then the absolute positions of the feature corners can be located. The robustness analysis of the algorithm with synthetic data and the experimental results of real face images are enclosed.

What is the probability of the Bayesian model, given the data?

The great advantage of Tenenbaum and Griffiths's model is that it incorporates both specific and general prior knowledge into category learning. Two phenomena are presented as supporting the detailed assumptions of this model. However, one phenomenon, effects of diversity, does not seem to require these assumptions, and the other phenomenon, effects of sample size, is not representative of most reported results. [Tenenbaum &amp; Griffiths]

Encoding effects of remindings

Understanding requires prior knowledge. Most theories and empirical work on understanding have focused on the use of general prior knowledge. In the experiments presented here, we examined the role of specific prior episodes in understanding. Subjects read ambiguous stories that had some superficial cue to make them think back to an earlier story. Experiment 1 shows that a superficial cue to an earlier story affects how the new story is interpreted. Experiment 2, in which a reading-time measure was used, provides evidence that this effect of the earlier story occurs during the initial understanding of the new story. We discuss the implications of these results for theories of understanding and the use of remindings in analogy.