Articles

Efficient instance retrieval of subgoals for subsumptive tabled evaluation of logic programs

Cruz F., Rocha R.

Theory and Practice of Logic Programming

2011

Abstract:

Tabled evaluation is an implementation technique that solves some problems of traditional Prolog systems in dealing with recursion and redundant computations. Most tabling engines determine if a tabled subgoal will produce or consume answers by using variant checks. A more refined method, named call subsumption, considers that a subgoal A will consume from a subgoal B if A is subsumed by (an instance of) B, thus allowing greater answer reuse. We recently developed an extension, called Retroactive Call Subsumption, that improves upon call subsumption by supporting bidirectional sharing of answers between subsumed/subsuming subgoals. In this paper, we present both an algorithm and an extension to the table space data structures to efficiently implement instance retrieval of subgoals for subsumptive tabled evaluation of logic programs. Experiments results using the YapTab tabling system show that our implementation performs quite well on some complex benchmarks and is robust enough to handle a large number of subgoals without performance degradation.

Efficient Marginalization of Discrete and Structured Latent Variables via Sparsity

Correia G., Niculae V., Aziz W., Martins A.

NeurIPS 2020

Abstract:

Training neural network models with discrete (categorical or structured) latent variables can be computationally challenging, due to the need for marginalization over large or combinatorial sets. To circumvent this issue, one typically resorts to sampling-based approximations of the true marginal, requiring noisy gradient estimators (e.g., score function estimator) or continuous relaxations with lower-variance reparameterized gradients (e.g., Gumbel-Softmax). In this paper, we propose a new training strategy which replaces these estimators by an exact yet efficient marginalization. To achieve this, we parameterize discrete distributions over latent assignments using differentiable sparse mappings: sparsemax and its structured counterparts. In effect, the support of these distributions is greatly reduced, which enables efficient marginalization. We report successful results in three tasks covering a range of latent variable modeling applications: a semisupervised deep generative model, a latent communication game, and a generative model with a bit-vector latent representation. In all cases, we obtain good performance while still achieving the practicality of sampling-based approximations.

Efficient methods for point matching with known camera orientation

Mota J.F.C., Aguiar P.M.Q.

Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

2010

Abstract:

The vast majority of methods that successfully recover 3D structure from 2D images hinge on a preliminary identification of corresponding feature points. When the images capture close views, e.g., in a video sequence, corresponding points can be found by using local pattern matching methods. However, to better constrain the 3D inference problem, the views must be far apart, leading to challenging point matching problems. In the recent past, researchers have then dealt with the combinatorial explosion that arises when searching among N! possible ways of matching N points. In this paper we overcome this search by making use of prior knowledge that is available in many situations: the orientation of the camera. This knowledge enables us to derive ( 2) algorithms to compute point correspondences. We prove that our approach computes the correct solution when dealing with noiseless data and derive an heuristic that results robust to the measurement noise and the uncertainty in prior knowledge. Although we model the camera using orthography, our experiments illustrate that our method is able to deal with violations, including the perspective effects of general real images.

Efficient Navigation in Learning Materials: An Empirical Study on the Linking Process

Mota P., Coheur L., Eskenazi M.

AIED 2018: Artificial Intelligence in Education

2018

Abstract:

We focus on the task of linking topically related segments in a collection of documents. In this scope, an existing corpus of learning materials was annotated with links between its segments. Using this corpus, we evaluate clustering, topic models, and graph-community detection algorithms in an unsupervised approach to the linking task. We propose several schemes to weight the word co-occurrence graph in order to discovery word communities, as well as a method for assigning segments to the discovered communities. Our experimental results indicate that the graph-community approach might BE more suitable for this task.

Efficient Privacy Preserving Distributed K-Means for Non-IID Data

Brandão A., Mendes R., Vilela J.

Advances in Intelligent Data Analysis XIX

2021

Abstract:

Privacy is becoming a crucial requirement in many machine learning systems. In this paper we introduce an efficient and secure distributed K-Means algorithm, that is robust to non-IID data. The base idea of our proposal consists in each client computing the K-Means algorithm locally, with a variable number of clusters. The server will use the resultant centroids to apply the K-Means algorithm again, discovering the global centroids. To maintain the client’s privacy, homomorphic encryption and secure aggregation is used in the process of learning the global centroids. This algorithm is efficient and reduces transmission costs, since only the local centroids are used to find the global centroids. In our experimental evaluation, we demonstrate that our strategy achieves a similar performance to the centralized version even in cases where the data follows an extreme non-IID form.

Effortless scanning of 3D object models by boundary aligning and stitching

Brandao S., Costeira J.P., Veloso M.

VISAPP 2014 - Proceedings of the 9th International Conference on Computer Vision Theory and Applications

2014

Abstract:

We contribute a novel algorithm for the digitation of complete 3D object models that requires little preparation effort from the user. Notably, the presented algorithm, Joint Alignment and Stitching of Non-Overlapping Meshes (JASNOM), completes 3D object models by aligning and stitching two 3D meshes by the boundaries and does not require any previous registration between them. JASNOM only requirement is the lack of overlap between meshes, which is simple to achieve in most man made object. JASNOM takes advantage that both meshes can only be connected by their boundary to reframe the alignment problem as a search of the best assignment between boundary vertices. To make the problem tractable, JASNOM reduces the search space considerably by imposing strong constraints on valid assignments that transform the original combinatorial problem into a discrete linear problem. By not requiring previous camera registration and by not depending on shape features, JASNOM contributions range from quick modeling of 3D objects to hole filling in meshes.

EGaIn‐Assisted Room‐Temperature Sintering of Silver Nanoparticles for Stretchable, Inkjet‐Printed, Thin‐Film Electronics

Tavakoli M., Malakooti M.H., Paisana H., Ohm Y., Marques D.G., Lopes P.A. , Piedade A.P., Almeida A.T., Majidi C.

Advanced Materials

2018

Abstract:

Coating inkjet‐printed traces of silver nanoparticle (AgNP) ink with a thin layer of eutectic gallium indium (EGaIn) increases the electrical conductivity by six‐orders of magnitude and significantly improves tolerance to tensile strain. This enhancement is achieved through a room‐temperature “sintering” process in which the liquid‐phase EGaIn alloy binds the AgNP particles (≈100 nm diameter) to form a continuous conductive trace. Ultrathin and hydrographically transferrable electronics are produced by printing traces with a composition of AgNP‐Ga‐In on a 5 µm‐thick temporary tattoo paper. The printed circuit is flexible enough to remain functional when deformed and can support strains above 80% with modest electromechanical coupling (gauge factor ≈1). These mechanically robust thin‐film circuits are well suited for transfer to highly curved and nondevelopable 3D surfaces as well as skin and other soft deformable substrates. In contrast to other stretchable tattoo‐like electronics, the low‐cost processing steps introduced here eliminate the need for cleanroom fabrication and instead requires only a commercial desktop printer. Most significantly, it enables functionalities like “electronic tattoos” and 3D hydrographic transfer that have not been previously reported with EGaIn or EGaIn‐based biphasic electronics.

Embedded Regularization For Classification Of Colposcopic Images

Albuquerque T., Cardoso J.

2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI)

2021

Abstract:

Cervical cancer ranks as the fourth most common cancer among females worldwide with roughly 528,000 new cases yearly. Significant progress in the realm of artificial intelligence particularly in neural networks and deep learning help physicians to diagnose cervical cancer more accurately. In this paper, we address a classification problem with the widely used VGG16 architecture. In addition to classification error, our model considers a regularization part during tuning of the weights, acting as prior knowledge of the colposcopic image. This embedded regularization approach, using a 2D Gaussian kernel, has enabled the model to learn which sections of the medical images are more crucial for the classification task. The experimental results show an improvement compared with standard transfer learning and multimodal approaches of cervical cancer classification in literature.

Emergent behavior in large scale networks

Santos A., Moura J.M.F.

Proceedings of the IEEE Conference on Decision and Control

2011

Abstract:

We apply mean field asymptotic analysis to explain the emergence of global behavior in large scale networks. The underlying motivating application is epidemics like computer virus spreading, for example, in wide campus local networks. We consider multiple classes of viruses, each type bearing their own statistical characterization – exogenous contamination, contagious propagation, and healing. The network state (distribution of nodes infected by each class in the network) is a jump Markov process, not necessarily reversible, making it a challenge to obtain its invariant distribution. By suitable renormalization, in the limit of a large network (number of nodes,) the macroscopic behavior of the network is described by the solution of a set of deterministic nonlinear differential equations (Riccati type.) We show that, under the heavy traffic assumption, the relevant underlying dynamics induces a coherent nontrivial metastable behavior in a macroscopic space-time scale: a slight imbalance on the effective spreading rate of one class over the others determines a significantly greater steady state predominance of this class over the others, regardless of the initial distribution.