A new transfer learning framework with application to model-agnostic multi-task learning

Learning from small number of examples is a challenging problem in machine learning. An effective way to improve the performance is through exploiting knowledge from other related tasks. Multi-task learning (MTL) is one such useful paradigm that aims to improve the performance through jointly modeling multiple related tasks. Although there exist numerous classification or regression models in machine learning literature, most of the MTL models are built around ridge or logistic regression. There exist some limited works, which propose multi-task extension of techniques such as support vector machine, Gaussian processes. However, all these MTL models are tied to specific classification or regression algorithms and there is no single MTL algorithm that can be used at a meta level for any given learning algorithm. Addressing this problem, we propose a generic, model-agnostic joint modeling framework that can take any classification or regression algorithm of a practitioner’s choice (standard or custom-built) and build its MTL variant. The key observation that drives our framework is that due to small number of examples, the estimates of task parameters are usually poor, and we show that this leads to an under-estimation of task relatedness between any two tasks with high probability. We derive an algorithm that brings the tasks closer to their true relatedness by improving the estimates of task parameters. This is achieved by appropriate sharing of data across tasks. We provide the detail theoretical underpinning of the algorithm. Through our experiments with both synthetic and real datasets, we demonstrate that the multi-task variants of several classifiers/regressors (logistic regression, support vector machine, K-nearest neighbor, Random Forest, ridge regression, support vector regression) convincingly outperform their single-task counterparts. We also show that the proposed model performs comparable or better than many state-of-the-art MTL and transfer learning baselines.

Multi-scale simulation using a hybrid mass-spring system and finite element model

Ply-scale finite element (FE) models are widely used to predict the performance of a composite structure based on material properties of individual plies. When simulating damage, these models neglect microscopic fracture processes which may have a significant effect on how a crack progresses within and between plies of a multidirectional laminate. To overcome this resolution limitation a multi-scale modelling technique is employed to simulate the effect micro-scale damage events have on the macro-scale response of a structure. The current paper discusses the development and validation of a hybrid mass-spring system and finite element modelling technique for multi-scale analysis. The model developed here is limited to elastic deformations; however, it is the first key step towards an efficient multi-scale damage model well suited to simulation of fracture in fibre reinforced composite materials. Various load cases have been simulated using the model developed here which show excellent accuracy compared to analytical and FE results. Future work is discussed, including extension of the model to incorporate damage modelling.

Acoustic emission study on corona discharge using FDTD model

This paper attempts to study the propagating characteristics of acoustic signals emitted from the breakdown of air using time domain numerical model. Acoustic emissions are produced by high voltage faults such as partial discharge and surface discharge. Study of such emissions has become popular among researchers because of the promising correlation between partial and surface discharges and its byproduct, acoustic signal emission. In this paper, propagation characteristics of acoustic signals are studied using finite difference time domain (FDTD) method. Multiple monitoring points are placed within a designated computation space at different distance away from a source.

Comparison of the strain distribution obtained from multi scale and conventional approaches to modelling extrusion

An investigation of the application of a multi scale CAFE model to prediction of the strain localization phenomena in industrial processes, such as extrusion, is presented in this work. Extrusion involves the formation of a strong strain localization zone, which influences the final product microstructure and may lead to a coarse grain layer close to the surface. Modelling of the shape of this zone and prediction of the strain magnitude will allow computer aided design of the extrusion process and optimisation of the technological parameters with respect to the microstructure and properties of the products. Thus, the particular objective of this work is comparison of the FE and CAFE predictions of strain localization in the shear zone area in extrusion. Advantages and disadvantages of the developed CAFE model are also discussed on the basis of the simulation results.

Concepts and dimensionality in modeling unsaturated water flow and solute transport

Many environmental studies require accurate simulation of water and solute fluxes in the unsaturated zone. This paper evaluates one- and multi-dimensional approaches for soil water flow as well as different spreading mechanisms to model solute behavior at different scales. For quantification of soil water fluxes,Richards equation has become the standard. Although current numerical codes show perfect water balances, the calculated soil water fluxes in case of head boundary conditions may depend largely on the method used for spatial averaging of the hydraulic conductivity. Atmospheric boundary conditions, especially in the case of phreatic groundwater levels fluctuating above and below a soil surface, require sophisticated solutions to ensure convergence. Concepts for flow in soils with macro pores and unstable wetting fronts are still in development. One-dimensional flow models are formulated to work with lumped parameters in order to account for the soil heterogeneity and preferential flow. They can be used at temporal and spatial scales that are of interest to water managers and policymakers. Multi-dimensional flow models are hampered by data and computation requirements.Their main strength is detailed analysis of typical multi-dimensional flow problems, including soil heterogeneity and preferential flow. Three physically based solute-transport concepts have been proposed to describe solute spreading during unsaturated flow: The stochastic-convective model (SCM), the convection-dispersion equation (CDE), and the fraction aladvection-dispersion equation (FADE). A less physical concept is the continuous-time random-walk process (CTRW). Of these, the SCM and the CDE are well established, and their strengths and weaknesses are identified. The FADE and the CTRW are more recent,and only a tentative strength weakness opportunity threat (SWOT)analysis can be presented at this time. We discuss the effect of the number of dimensions in a numerical model and the spacing between model nodes on solute spreading and the values of the solute-spreading parameters. In order to meet the increasing complexity of environmental problems, two approaches of model combination are used: Model integration and model coupling. Amain drawback of model integration is the complexity of there sulting code. Model coupling requires a systematic physical domain and model communication analysis. The setup and maintenance of a hydrologic framework for model coupling requires substantial resources, but on the other hand, contributions can be made by many research groups.

Application of the strain localization CAFE model to investigate extrusion with various die shapes

Multi scale CAFE model for the prediction of initiation and propagation of the micro shear bands and shear bands in metallic materials subjected to plastic deformation is presented. The CAFE approach is the combination of the Cellular Automata (CA) and the Finite Element (FE) methods. The application of the developed CAFE model to analyze material flow during extrusion is the objective of the present work. The proposed CAFE approach is applied in this work to simulation of the extrusion with flat face and convex dies and to investigate differences in the material flow. The initial FE meshes with the set of the CA point are generated for the numerical tests and the results of the metal flow predicted by the CAFE method are presented in the paper.

Architectural information

Through a study of architectural information as a complex system, a theoretical model relating critical information elements and project knowledge is proposed. The Project Specific Information model articulates distinct information integrities (intellectual, contextual, structural and spatial) that support the knowledge-based decision-making processes controlling information quality in architectural projects.

Matrices representation of multi soft-sets and its application

In previous paper, we introduced a concept of multi-soft sets and used it for finding reducts. However, the comparison of the proposed reduct has not been presented yet, especially with rough-set based reduct. In this paper, we present matrices representation of multi-soft sets. We define AND and OR operations on a collection of such matrices and apply it for finding reducts and core of attributes in a multi-valued information system. Finally, we prove that our proposed technique for reduct is equivalent to Pawlak’s rough reduct.

Volume confinement induced microstructural transitions and property enhancements of supramolecular soft materials

The rheological properties of supramolecular soft functional materials are determined by the networks within the materials. This research reveals for the first time that the volume confinement during the formation of supramolecular soft functional materials will exert a significant impact on the rheological properties of the materials. A class of small molecular organogels formed by the gelation of N-lauroyl-L-glutamic acid din-butylamide (GP-1) in ethylene glycol (EG) and propylene glycol (PG) solutions were adopted as model systems for this study. It follows that within a confined space, the elasticity of the gel can be enhanced more than 15 times compared with those under un-restricted conditions. According to our optical microscopy observations and rheological measurements, this drastic enhancement is caused by the structural transition from a multi-domain network system to a single network system once the average size of the fiber network of a given material reaches the lowest dimension of the system. The understanding acquired from this work will provide a novel strategy to manipulate the network structure of soft materials, and exert a direct impact on the micro-engineering of such supramolecular materials in micro and nano scales.

M-Dimension : multi-characteristics based routing protocol in human associated delay-tolerant networks with improved performance over one dimensional classic models

Human associated delay-tolerant network (HDTN) is a new delay-tolerant network where mobile devices are associated with humans. It can be viewed from both their geographic and social dimensions. The combination of these different dimensions can enable us to more accurately comprehend a delay-tolerant network and consequently use this multi-dimensional information to improve overall network efficiency. Alongside the geographic dimension of the network which is concerned with geographic topology of routing, social dimensions such as social hierarchy can be used to guide the routing message to improve not only the routing efficiency for individual nodes, but also efficiency for the entire network.

We propose a multi-dimensional routing protocol (M-Dimension) for the human associated delay-tolerant network which uses the local information derived from multiple dimensions to identify a mobile node more accurately. Each dimension has a weight factor and is organized by the Distance Function to select an intermediary and applies multi-cast routing. We compare M-Dimension to existing benchmark routing protocols using the MIT Reality Dataset, a well-known benchmark dataset based on a human associated mobile network trace file. The results of our simulations show that M-Dimension has a significant increase in the average success ratio and is very competitive when End-to-End Delay of packet delivery is used in comparison to other multi-cast DTN routing protocols.

A non-linear programming-based similarity reasoning scheme for modelling of monotonicity-preserving multi-input fuzzy inference systems

In this paper, the zero-order Sugeno Fuzzy Inference System (FIS) that preserves the monotonicity property is studied. The sufficient conditions for the zero-order Sugeno FIS model to satisfy the monotonicity property are exploited as a set of useful governing equations to facilitate the FIS modelling process. The sufficient conditions suggest a fuzzy partition (at the rule antecedent part) and a monotonically-ordered rule base (at the rule consequent part) that can preserve the monotonicity property. The investigation focuses on the use of two Similarity Reasoning (SR)-based methods, i.e., Analogical Reasoning (AR) and Fuzzy Rule Interpolation (FRI), to deduce each conclusion separately. It is shown that AR and FRI may not be a direct solution to modelling of a multi-input FIS model that fulfils the monotonicity property, owing to the difficulty in getting a set of monotonically-ordered conclusions. As such, a Non-Linear Programming (NLP)-based SR scheme for constructing a monotonicity-preserving multi-input FIS model is proposed. In the proposed scheme, AR or FRI is first used to predict the rule conclusion of each observation. Then, a search algorithm is adopted to look for a set of consequents with minimized root means square errors as compared with the predicted conclusions. A constraint imposed by the sufficient conditions is also included in the search process. Applicability of the proposed scheme to undertaking fuzzy Failure Mode and Effect Analysis (FMEA) tasks is demonstrated. The results indicate that the proposed NLP-based SR scheme is useful for preserving the monotonicity property for building a multi-input FIS model with an incomplete rule base.

Integrated construction supply chain design and delivery solutions

This study describes the development of a decision framework to support multi-disciplinary information and knowledge management model which focuses on integrated design and delivery solutions for all construction supply chain actors. The framework was developed within the context of two national information technology research projects in Australia. The first study used diffusion theory to explain the barriers and enablers to future adoption of advanced information technology solutions such as building information modelling (BIM). A grounded theory methodology was deployed and a pathways model for innovative information technology diffusion accommodating diverse patterns of adoption and different levels of expertize was developed. The second study built on the findings of the first study but specifically focussed on innovators, early and late adopters of BIM and the development of a decision framework towards advanced collaborative platform solutions. This study summarizes the empirical results of the previous studies. The core of the decision framework is the creation, use and ownership of building information sub-models and integrated models. The decision framework relies on holistic collaborative design management. Design expertise is diffused and can be found in various locations along the construction supply chain within project teams. A wide definition of design is considered from conceptual to developed to detailed design. The recent development to the decision model offers much potential as the early upstream decisions are often made in a creative, collaborative and uncertain environment. However, decision making needs to balance both a reductionist and exploratory creative empowerment approach. Shared team expertise and competency and team mental models are explored as a fundamental requirement to collaborative BIM. New skills in interdisciplinarity are discussed as an implication of future construction industry collaborative platforms.

Adaptable, model-driven security engineering for SaaS cloud-based applications

Software-as-a-service (SaaS) multi-tenancy in cloud-based applications helps service providers to save cost, improve resource utilization, and reduce service customization and maintenance time. This is achieved by sharing of resources and service instances among multiple "tenants" of the cloud-hosted application. However, supporting multi-tenancy adds more complexity to SaaS applications required capabilities. Security is one of these key requirements that must be addressed when engineering multi-tenant SaaS applications. The sharing of resources among tenants - i.e. multi-tenancy - increases tenants' concerns about the security of their cloud-hosted assets. Compounding this, existing traditional security engineering approaches do not fit well with the multi-tenancy application model where tenants and their security requirements often emerge after the applications and services were first developed. The resultant applications do not usually support diverse security capabilities based on different tenants' needs, some of which may change at run-time i.e. after cloud application deployment. We introduce a novel model-driven security engineering approach for multi-tenant, cloud-hosted SaaS applications. Our approach is based on externalizing security from the underlying SaaS application, allowing both application/service and security to evolve at runtime. Multiple security sets can be enforced on the same application instance based on different tenants' security requirements. We use abstract models to capture service provider and multiple tenants' security requirements and then generate security integration and configurations at runtime. We use dependency injection and dynamic weaving via Aspect-Oriented Programming (AOP) to integrate security within critical application/service entities at runtime. We explain our approach, architecture and implementation details, discuss a usage example, and present an evaluation of our approach on a set of open source web applications.

What shall i share and with whom? A multi-task learning formulation using multi-faceted task relationships

Multi-task learning is a learning paradigm that improves the performance of "related" tasks through their joint learning. To do this each task answers the question "Which other task should I share with"? This task relatedness can be complex - a task may be related to one set of tasks based on one subset of features and to other tasks based on other subsets. Existing multi-task learning methods do not explicitly model this reality, learning a single-faceted task relationship over all the features. This degrades performance by forcing a task to become similar to other tasks even on their unrelated features. Addressing this gap, we propose a novel multi-task learning model that leams multi-faceted task relationship, allowing tasks to collaborate differentially on different feature subsets. This is achieved by simultaneously learning a low dimensional sub-space for task parameters and inducing task groups over each latent subspace basis using a novel combination of L1 and pairwise L∞ norms. Further, our model can induce grouping across both positively and negatively related tasks, which helps towards exploiting knowledge from all types of related tasks. We validate our model on two synthetic and five real datasets, and show significant performance improvements over several state-of-the-art multi-task learning techniques. Thus our model effectively answers for each task: What shall I share and with whom?

Possibilities to include in numerical simulation stochastic and discontinuous phenomena occurring in material

Simulation of materials processing has to face new difficulties regarding proper description of various discontinuous and stochastic phenomena occurring in materials. Commonly used rheological models based on differential equations treat material as continuum and are unable to describe properly several important phenomena. That is the reason for ongoing search for alternative models, which can account for non-continuous structure of the materials and for the fact, that various phenomena in the materials occur in different scales from nano to mezo. Accounting for the stochastic character of some phenomena is an additional challenge. One of the solutions may be the coupled Cellular Automata (CA) – Finite Element (FE) multi scale model. A detailed discussion about the advantages given by the developed multi scale CAFE model for strain localization phenomena in contrast to capabilities provided by the conventional FE approaches is a subject of this work. Results obtained from the CAFE model are supported by the experimental observations showing influence of many discontinuities existing in the real material on macroscopic response. An immense capabilities of the CAFE approach in comparison to limitations of the FE method for modeling of real material behavior is are shown this work as well.