Given a convex and differentiable objective $Q(\M)$ for a real, symmetric matrix $\M$ in the positive definite (PD) cone---used to compute Mahalanobis distances---we propose a fast general metric learning framework that is entirely projection-free. We first assume that $\M$ resides in a space $\cS$ of generalized graph Laplacian matrices (graph metric matrices) corresponding to balanced signed graphs. Unlike low-rank metric matrices common in the literature, $\cS$ includes the important diagonal-only matrices as a special case. The key theorem to circumvent full eigen-decomposition and enable fast metric matrix optimization is Gershgorin disc alignment (GDA): given graph metric matrix $\M \in \cS$ and diagonal matrix $\S$, where $S_{ii} = 1/v_i$ and $\v$ is the first eigenvector of $\M$, we prove that Gershgorin disc left-ends of similar transform $\B = \S \M \S^{-1}$ are perfectly aligned at the smallest eigenvalue $\lambda_{\min}$. Using this theorem, we replace the PD cone constraint in the metric learning problem with tightest possible linear constraints per iteration, so that the alternating optimization of the diagonal / off-diagonal terms in $\M$ can be solved efficiently as linear programs via Frank-Wolfe iterations. We update $\v$ using Locally Optimal Block Preconditioned Conjugate Gradient (LOBPCG) with warm start as matrix entries in $\M$ are optimized successively. Experiments show that our graph metric optimization is significantly faster than cone-projection methods, and produces competitive binary classification performance.
Modern neural networks are often regarded as complex black-box functions whose behavior is difficult to understand owing to their nonlinear dependence on the data and the nonconvexity in their loss landscapes. In this work, we show that these common perceptions can be completely false in the early phase of learning. In particular, we formally prove that, for a class of well-behaved input distributions, the early-time learning dynamics of a two-layer fully-connected neural network can be mimicked by training a simple linear model on the inputs. We additionally argue that this surprising simplicity can persist in networks with more layers and with convolutional architecture, which we verify empirically. Key to our analysis is to bound the spectral norm of the difference between the Neural Tangent Kernel (NTK) at initialization and an affine transform of the data kernel; however, unlike many previous results utilizing the NTK, we do not require the network to have disproportionately large width, and the network is allowed to escape the kernel regime later in training.
Particle filtering is a popular method for inferring latent states in stochastic dynamical systems, whose theoretical properties have been well studied in machine learning and statistics communities. In sequential decision-making problems, e.g., partially observed Markov decision processes (POMDPs), oftentimes the inferred latent state is further used for planning at each step. This paper initiates a rigorous study on the efficiency of particle filtering for sequential planning, and gives the first particle complexity bounds. Though errors in past actions may affect the future, we are able to bound the number of particles needed so that the long-run reward of the policy based on particle filtering is close to that based on exact inference. In particular, we show that, in stable systems, polynomially many particles suffice. Key in our analysis is a coupling of the ideal sequence based on the exact planning and the sequence generated by approximate planning based on particle filtering. We believe this technique can be useful in other sequential decision-making problems.
Learning knowledge graph (KG) embeddings has received increasing attention in recent years. Most embedding models in literature interpret relations as linear or bilinear mapping functions to operate on entity embeddings. However, we find that such relation-level modeling cannot capture the diverse relational structures of KGs well. In this paper, we propose a novel edge-centric embedding model TransEdge, which contextualizes relation representations in terms of specific head-tail entity pairs. We refer to such contextualized representations of a relation as edge embeddings and interpret them as translations between entity embeddings. TransEdge achieves promising performance on different prediction tasks. Our experiments on benchmark datasets indicate that it obtains the state-of-the-art results on embedding-based entity alignment. We also show that TransEdge is complementary with conventional entity alignment methods. Moreover, it shows very competitive performance on link prediction.
The prevalence of accessible depth sensing and 3D laser scanning techniques has enabled the convenient acquisition of 3D dynamic point clouds, which provide efficient representation of arbitrarily-shaped objects in motion. Nevertheless, dynamic point clouds are often perturbed by noise due to hardware, software or other causes. While a plethora of methods have been proposed for static point cloud denoising, few efforts are made for the denoising of dynamic point clouds with varying number of irregularly-sampled points in each frame. In this paper, we represent dynamic point clouds naturally on graphs and address the denoising problem by inferring the underlying graph via spatio-temporal graph learning, exploiting both the intra-frame similarity and inter-frame consistency. Firstly, assuming the availability of a relevant feature vector per node, we pose spatial-temporal graph learning as optimizing a Mahalanobis distance metric M, which is formulated as the minimization of graph Laplacian regularizer. Secondly, to ease the optimization of the symmetric and positive definite metric matrix M, we decompose it into M = R'*R and solve R instead via proximal gradient. Finally, based on the spatial-temporal graph learning, we formulate dynamic point cloud denoising as the joint optimization of the desired point cloud and underlying spatio-temporal graph, which leverages both intra-frame affinities and inter-frame consistency and is solved via alternating minimization. Experimental results show that the proposed method significantly outperforms independent denoising of each frame from state-of-the-art static point cloud denoising approaches.
In the active research area of employing embedding models for knowledge graph completion, particularly for the task of link prediction, most prior studies used two benchmark datasets FB15k and WN18 in evaluating such models. Most triples in these and other datasets in such studies belong to reverse and duplicate relations which exhibit high data redundancy due to semantic duplication, correlation or data incompleteness. This is a case of excessive data leakage---a model is trained using features that otherwise would not be available when the model needs to be applied for real prediction. There are also Cartesian product relations for which every triple formed by the Cartesian product of applicable subjects and objects is a true fact. Link prediction on the aforementioned relations is easy and can be achieved with even better accuracy using straightforward rules instead of sophisticated embedding models. A more fundamental defect of these models is that the link prediction scenario, given such data, is non-existent in the real-world. This paper is the first systematic study with the main objective of assessing the true effectiveness of embedding models when the unrealistic triples are removed. Our experiment results show these models are much less accurate than what we used to perceive. Their poor accuracy renders link prediction a task without truly effective automated solution. Hence, we call for re-investigation of possible effective approaches.
Entity alignment seeks to find entities in different knowledge graphs (KGs) that refer to the same real-world object. Recent advancement in KG embedding impels the advent of embedding-based entity alignment, which encodes entities in a continuous embedding space and measures entity similarities based on the learned embeddings. In this paper, we conduct a comprehensive experimental study of this emerging field. This study surveys 23 recent embedding-based entity alignment approaches and categorizes them based on their techniques and characteristics. We further observe that current approaches use different datasets in evaluation, and the degree distributions of entities in these datasets are inconsistent with real KGs. Hence, we propose a new KG sampling algorithm, with which we generate a set of dedicated benchmark datasets with various heterogeneity and distributions for a realistic evaluation. This study also produces an open-source library, which includes 12 representative embedding-based entity alignment approaches. We extensively evaluate these approaches on the generated datasets, to understand their strengths and limitations. Additionally, for several directions that have not been explored in current approaches, we perform exploratory experiments and report our preliminary findings for future studies. The benchmark datasets, open-source library and experimental results are all accessible online and will be duly maintained.