Literatura académica sobre el tema "Partially-Separable Structure"

Objective: Previous studies in children with ADHD identified two partially separable familial factors underlying cognitive dysfunction, but evidence in adolescents and adults is lacking. Here, we investigate the etiological structure of cognitive-neurophysiological impairments in ADHD in adolescents and young adults. Method: Factor analyses and multivariate familial models were run in 356 participants from ADHD and control sibling pairs aged 11 to 27 years on data on IQ, digit span forward (DSF) and backward (DSB), and cognitive-performance and event-related potential (ERP) measures from three cognitive tasks. Results: Three familial factors (cF1-3), showing substantial familial overlap with ADHD, captured the familial covariation of ADHD with nine cognitive-ERP measures. cF1 loaded on IQ, mean reaction time (MRT), and reaction-time variability (RTV); cF2 on DSF and DSB; and cF3 on number of errors and ERPs of inhibition and error processing. Conclusion: These results identify three partially separable etiological pathways leading to cognitive-neurophysiological impairments in adolescent and adult ADHD.

A structured version of derivative-free random pattern search optimization algorithms is introduced, which is able to exploit coordinate partially separable structure (typically associated with sparsity) often present in unconstrained and bound-constrained optimization problems. This technique improves performance by orders of magnitude and makes it possible to solve large problems that otherwise are totally intractable by other derivative-free methods. A library of interpolation-based modelling tools is also described, which can be associated with the structured or unstructured versions of the initial pattern search algorithm. The use of the library further enhances performance, especially when associated with structure. The significant gains in performance associated with these two techniques are illustrated using a new freely-available release of the Brute Force Optimizer (BFO) package firstly introduced in [Porcelli and Toint 2017 ], which incorporates them. An interesting conclusion of the numerical results presented is that providing global structural information on a problem can result in significantly less evaluations of the objective function than attempting to building local Taylor-like models.

AbstractOne partially ordered set, Q, is a Tukey quotient of another, P, if there is a map (a Tukey quotient) $\phi :P \to Q$ carrying cofinal sets of P to cofinal sets of Q. Two partial orders which are mutual Tukey quotients of each other are said to be Tukey equivalent. Let ${\cal D}_{\rm{}} $ be the partially ordered set of Tukey equivalence classes of directed sets of size $ \le {\rm{}}$. It is shown that ${\cal D}_{\rm{}} $ contains an antichain of size $2^{\rm{}} $, and so has size $2^{\rm{}} $. The elements of the antichain are of the form ${\cal K}\left( M \right)$, the set of compact subsets of a separable metrizable space M, ordered by inclusion. The order structure of such ${\cal K}\left( M \right)$’s under Tukey quotients is investigated. Relative Tukey quotients are introduced. Applications are given to function spaces and to the complexity of weakly countably determined Banach spaces and Gul’ko compacta.

Background: Some individuals with autism find it challenging to use and understand language in conversation, despite having good abilities in core aspects of language such as grammar and vocabulary. This suggests that pragmatic skills (such as understanding implied meanings in conversation) are separable from core language skills. However, it has been surprisingly difficult to demonstrate this dissociation in the general population. We propose that this may be because prior studies have used tasks in which different aspects of language are confounded. Methods: The present study used novel language tasks and factor analysis to test whether pragmatic language skills are separable from core language skills. 120 adult participants were recruited online to complete a 7-task battery, including a test assessing comprehension of conversational implicature. Results: In confirmatory analysis of a preregistered model, we compared whether the data showed better fit to a two-factor structure (including a pragmatic conversation comprehension and core language factor) or a simpler one-factor structure (comprising a general language factor). The two-factor model showed significantly better fit. Conclusions: This study supports the view that interpreting context-dependent conversational meaning is partially distinct from core language skill. This has implications for understanding the pragmatic language impairments reported in autism.

Background: Some individuals with autism find it challenging to use and understand language in conversation, despite having good abilities in core aspects of language such as grammar and vocabulary. This suggests that pragmatic skills (such as understanding implied meanings in conversation) are separable from core language skills. However, it has been surprisingly difficult to demonstrate this dissociation in the general population. We propose that this may be because prior studies have used tasks in which different aspects of language are confounded. Methods: The present study used novel language tasks and factor analysis to test whether pragmatic understanding of implied meaning, as part of a broader domain involving social understanding, is separable from core language skills. 120 adult participants were recruited online to complete a 7-task battery, including a test assessing comprehension of conversational implicature. Results: In confirmatory analysis of a preregistered model, we compared whether the data showed better fit to a two-factor structure (including a “social understanding” and “core language” factor) or a simpler one-factor structure (comprising a general factor). The two-factor model showed significantly better fit. Conclusions: This study supports the view that interpreting context-dependent conversational meaning is partially distinct from core language skills. This has implications for understanding the pragmatic language impairments reported in autism.

Background: Some individuals with autism find it challenging to use and understand language in conversation, despite having good abilities in core aspects of language such as grammar and vocabulary. This suggests that pragmatic skills (such as understanding implied meanings in conversation) are separable from core language skills. However, it has been surprisingly difficult to demonstrate this dissociation in the general population. We propose that this may be because prior studies have used tasks in which different aspects of language are confounded. Methods: The present study used novel language tasks and factor analysis to test whether pragmatic understanding of implied meaning, as part of a broader domain involving social understanding, is separable from core language skills. 120 adult participants were recruited online to complete a 7-task battery, including a test assessing comprehension of conversational implicature. Results: In confirmatory analysis of a preregistered model, we compared whether the data showed better fit to a two-factor structure (including a “social understanding” and “core language” factor) or a simpler one-factor structure (comprising a general factor). The two-factor model showed significantly better fit. Conclusions: This study supports the view that interpreting context-dependent conversational meaning is partially distinct from core language skills. This has implications for understanding the pragmatic language impairments reported in autism.

Factor analyses suggest that impulsivity traits that capture tendencies to act prematurely or take risks tap partially distinct constructs. We applied genomic structure equation modeling to evaluate the genetic factor structure of two well-established impulsivity questionnaires, using published statistics from genome-wide association studies of up to 22,861 participants. We also tested the hypotheses that delay discounting would be genetically separable from other impulsivity factors and that emotionally triggered facets of impulsivity (urgency) would be those most strongly genetically correlated with an internalizing latent factor. A five-factor model best fitted the impulsivity data. Delay discounting was genetically distinct from these five factors. As expected, the two urgency subscales were most strongly related to an internalizing-psychopathology latent factor. These findings provide empirical genetic evidence that impulsivity can be broken down into distinct categories of differential relevance for internalizing psychopathology. They also demonstrate how measured genetic markers can be used to inform theories of psychology and personality.

At present, ResNet and DenseNet have achieved significant performance gains in the field of finger-vein biometric recognition, which is partially attributed to the dominant design of cross-layer skip connection. In this manner, features from multiple layers can be effectively aggregated to provide sufficient discriminant representation. Nevertheless, an over-dense connection pattern may induce channel expansion of feature maps and excessive memory consumption. To address these issues, we proposed a low memory overhead and fairly lightweight network architecture for finger-vein recognition. The core components of the proposed network are a sequence of sparsified densely connected blocks with symmetric structure. In each block, a novel connection cropping strategy is adopted to balance the channel ratio of input/output feature maps. Beyond this, to facilitate smaller model volume and faster convergence, we substitute the standard convolutional kernels with separable convolutional kernels and introduce a robust loss metric that is defined on the geodesic distance of angular space. Our proposed sparsified densely connected network with separable convolution (hereinafter dubbed ‘SC-SDCN’) has been tested on two benchmark finger-vein datasets, including the Multimedia Lab of Chonbuk National University (MMCBNU)and Finger Vein of Universiti Sains Malaysia (FV-USM), and the advantages of our SC-SDCN can be evident from the experimental results. Specifically, an equal error rate (EER) of 0.01% and an accuracy of 99.98% are obtained on the MMCBNU dataset, and an EER of 0.45% and an accuracy of 99.74% are obtained on the FV-USM dataset.

Cette thèse est centrée autour de l'utilisation de la structure partiellement séparable pour l'optimisation sans contraintes, notamment pour les méthodes quasi-Newton et l'entraînement des réseaux de neurones.Une fonction partiellement séparable somme des fonctions éléments, chacune de dimension inférieure au problème total.Ainsi, le Hessian peut être agrégé en approximant séparément le hessien de chaque fonction élément par une matrice dense.Ces méthodes quasi-Newton partitionnées sont applicables aux problèmes de grandes dimensions et conservent la structure creuse du hessien, contrairement à une méthode quasi-Newton à mémoire limitée.En pratique, ces méthodes réalisent moins d'itérations qu'une méthode quasi-Newton à mémoire limitée, et sont parallélisables en distribuant les calculs reliés aux fonctions éléments.La revue de littérature complète sur le sujet a cependant révélé certaines limitations, particulièrement lorsque la dimension des fonctions éléments est large.De plus, l'unique logiciel libre d'optimisation exploitant la structure partiellement-séparable est inutilisables pour utilisateur non expérimenté, laissant pour seul choix le recourt à des logiciels commerciaux.Dans cette thèse, deux solutions sont proposées pour palier à ces lacunes ainsi qu'une application des concepts d'optimisation partiellement-séparable à l'apprentissage supervisé d'un réseau de neurones.La première contribution est une suite logicielle permettant notamment de détecter automatiquement la structure partiellement-séparable d'un problème, c'est-à-dire la récupération de chaque fonction élément de dimension réduite.Suite à cela, les structures de données partitionnées nécessaires à la mémorisation des dérivées, ou leurs approximations, sont allouées et permettent de définir des méthodes quasi-Newton partitionnées.L'ensemble est intégré à l'écosystème "JuliaSmoothOptimizers", regroupant de nombreux outils pour l'optimisation lisse, dont des algorithmes d'optimisation pouvant exploiter la séparabilité partielle détectée.La seconde contribution remplace l'approximation d'un hessien élément par une matrice dense par un opérateur linéaire quasi-Newton à mémoire limitée.Ce faisant, le coût mémoire de l'approximation totale du hessien n'est plus relié quadratiquement à la dimension des fonctions éléments.Une méthode quasi-Newton partitionnée à mémoire limitée est alors applicable lorsque les fonctions éléments sont de grandes tailles.Chaque méthode quasi-Newton partitionnée à mémoire limitée possède une preuve convergence globale.De plus, les résultats numériques montrent que ces méthodes surpassent les méthodes quasi-Newton partitionnées ou à mémoire limitées lorsque les éléments sont de grandes tailles.La dernière contribution étudie l'exploitation de la structure partiellement-séparable lors l'entraînement supervisé d'un réseau de neurones.Le problème d'optimisation lié à l'entraînement n'est généralement pas partiellement-séparable.Dès lors, une fonction de perte partiellement-séparable ainsi qu'une architecture de réseau partitionnée sont introduites afin de rendre l'entraînement partiellement-séparable.Les résultats numériques combinant ces deux apports sont compétitifs avec des architectures et des fonctions de perte standards selon les méthodes d'entraînement de l'état de l'art.De surcroît, cette combinaison permet un schéma de parallélisation supplémentaire aux méthodes déjà existantes pour l'apprentissage supervisé.En effet, les calculs de chaque fonction de perte élément peuvent être distribués à un travailleur nécessitant seulement une fraction du réseau de neurones pour opérer.Finalement, un entraînement quasi-Newton partitionnée à mémoire limitée est proposé.Cet entrainement est montré empiriquement comme compétitif avec les méthodes d'entraînement de l'état de l'art
This thesis studies and improves the use of the partially-separable structure for unconstrained optimization, particularly for quasi-Newton methods and training neural networks.A partially-separable function is the sum of element functions, each of lower dimension than the total problem.Thus, the Hessian can be aggregated by separately approximating the Hessian of each element function with a dense matrix.These partitioned quasi-Newton methods are applicable to high-dimensional problems and maintain the sparse structure of the Hessian, unlike a limited-memory quasi-Newton method.In practice, these methods require fewer iterations than a limited-memory quasi-Newton method and are parallelizable by distributing computations related to the element functions.However, a comprehensive literature review on the subject has revealed some limitations, particularly when the dimension of the element functions is large.Additionally, the only open-source optimization software exploiting the partially-separable structure is unusable for inexperienced users, leaving only commercial software as an option.In this thesis, two solutions are proposed to address these shortcomings, along with an application of partially-separable optimization concepts to supervised learning of a neural network.The first contribution is a software suite based on an automatic detection of the partially-separable structure of a problem, i.e., retrieves each reduced-dimensional element function.Following this, partitioned data structures necessary for storing derivatives, or their approximations, are allocated and used to define partitioned quasi-Newton optimization methods.The entire suite is integrated into the "JuliaSmoothOptimizers" ecosystem, which gathers numerous tools for smooth optimization, including optimization algorithms that can therefore exploit the detected partial separability.The second contribution replaces the approximation of an element Hessian by a dense matrix with a limited-memory quasi-Newton linear operator.As a result, the memory cost of the total Hessian approximation is no longer quadratically related to the dimension of the element functions.A limited-memory partitioned quasi-Newton method is then applicable when the element functions are large.Each limited-memory partitioned quasi-Newton method has a proof of global convergence.Additionally, numerical results show that these methods outperform partitioned or limited-memory quasi-Newton methods when the elements are large.The final contribution examines the exploitation of the partially-separable structure during supervised training of a neural network.The optimization problem associated with training is generally not partially-separable.Therefore, a partially-separable loss function and a partitioned network architecture are introduced to make the training partially-separable.Numerical results combining these two contributions are competitive with standard architectures and loss functions according to state-of-the-art training methods.Moreover, this combination produces an additional parallelization scheme to existing methods for supervised learning.Indeed, the calculations of each element loss function can be distributed to a worker requiring only a fraction of the neural network to operate.Finally, a limited-memory partitioned quasi-Newton training is proposed.This training is empirically shown to be competitive with state-of-the-art training methods

Abstract The problem of determining the optimal sizing design of truss structures is considered. An augmented Lagrangian optimization algorithm which uses a quadratic penalty term is formulated. The implementation uses a first-order Lagrange multiplier update and a strategy for progressively increasing the accuracy with which the bound constrained minimizations are performed. The allowed constraint violation is also progressively decreased but at a slower rate so as to prevent ill-conditioning due to large penalty values. Individual constraint penalties are used and only the penalties of the worst violated constraints are increased. The scheme is globally convergent. The bound constrained minimizations are performed using the SBMIN algorithm where a sophisticated trust-region strategy is employed. The Hessian of the augmented Lagrangian function is approximated using partitioned secant updating. Each function contributing to the Lagrangian is individually approximated by a secant update and the augmented Lagrangian Hessian is formed by appropriate accumulation. The performance of the algorithm is evaluated for a number of different secant updates on standard explicit and truss sizing optimization problems. The results show the formulation to be superior to other implementations of augmented Lagrangian methods reported in the literature and that, under certain conditions, the method approaches the performance of the state-of-the-art SQP and SAM methods. Of the secant updates, the symmetric rank one update is superior to the other updates including the BFGS scheme. It is suggested that the individual function, secant updating employed may be usefully applied in contexts where structural analysis and optimization are performed simultaneously, as in the simultaneous analysis and design method. In such cases the functions are partially separable and the associated Hessians are of low rank.