We are interested in an evaluation methodology for molecular optimization. Given a sample of molecules and their properties of our interest, we wish not only to train a generator of molecules optimized with respect to a target property but also to evaluate its performance accurately. A common practice is to train a predictor of the target property using the sample and apply it to both training and evaluating the generator. However, little is known about its statistical properties, and thus, we are not certain about whether this performance estimate is reliable or not. We theoretically investigate this evaluation methodology and show that it potentially suffers from two biases; one is due to misspecification of the predictor and the other to reusing the same finite sample for training and evaluation. We discuss bias reduction methods for each of the biases, and empirically investigate their effectiveness.

Current domain-independent, classical planners require symbolic models of the problem domain and instance as input, resulting in a knowledge acquisition bottleneck. Meanwhile, although deep learning has achieved significant success in many fields, the knowledge is encoded in a subsymbolic representation which is incompatible with symbolic systems such as planners. We propose Latplan, an unsupervised architecture combining deep learning and classical planning. Given only an unlabeled set of image pairs showing a subset of transitions allowed in the environment (training inputs), Latplan learns a complete propositional PDDL action model of the environment. Later, when a pair of images representing the initial and the goal states (planning inputs) is given, Latplan finds a plan to the goal state in a symbolic latent space and returns a visualized plan execution. We evaluate Latplan using image-based versions of 6 planning domains: 8-puzzle, 15-Puzzle, Blocksworld, Sokoban and Two variations of LightsOut.

This paper is concerned about a learning algorithm for a probabilistic model of spiking neural networks (SNNs). Jimenez Rezende & Gerstner (2014) proposed a stochastic variational inference algorithm to train SNNs with hidden neurons. The algorithm updates the variational distribution using the score function gradient estimator, whose high variance often impedes the whole learning algorithm. This paper presents an alternative gradient estimator for SNNs based on the path-wise gradient estimator. The main technical difficulty is a lack of a general method to differentiate a realization of an arbitrary point process, which is necessary to derive the path-wise gradient estimator. We develop a differentiable point process, which is the technical highlight of this paper, and apply it to derive the path-wise gradient estimator for SNNs. We investigate the effectiveness of our gradient estimator through numerical simulation.

We approach the time-series forecasting problem in the presence of concept drift by automatic learning rate tuning of stochastic gradient descent (SGD). The SGD-based approach is preferable to other concept drift algorithms in that it can be applied to any model and it can keep learning efficiently whilst predicting online. Among a number of SGD algorithms, the variance-based SGD (vSGD) can successfully handle concept drift by automatic learning rate tuning, which is reduced to an adaptive mean estimation problem. However, its performance is still limited because of its heuristic mean estimator. In this paper, we present a concept-drift-aware stochastic gradient descent (Cogra), equipped with more theoretically-sound mean estimator called sequential mean tracker (SMT). Our key contribution is that we define a goodness criterion for the mean estimators; SMT is designed to be optimal according to this criterion. As a result of comprehensive experiments, we find that (i) our SMT can estimate the mean better than vSGD’s estimator in the presence of concept drift, and (ii) in terms of predictive performance, Cogra reduces the predictive loss by 16–67% for real-world datasets, indicating that SMT improves the prediction accuracy significantly.

Molecular optimization aims to discover novel molecules with desirable properties, and its two fundamental challenges are: (i) it is not trivial to generate valid molecules in a controllable way due to hard chemical constraints such as the valency conditions, and (ii) it is often costly to evaluate a property of a novel molecule, and therefore, the number of property evaluations is limited. These challenges are to some extent alleviated by a combination of a variational autoencoder (VAE) and Bayesian optimization (BO), where VAE converts a molecule into/from its latent continuous vector, and BO optimizes a latent continuous vector (and its corresponding molecule) within a limited number of property evaluations. While the most recent work, for the first time, achieved 100% validity, its architecture is rather complex due to auxiliary neural networks other than VAE, making it difficult to train. This paper presents a molecular hypergraph grammar variational autoencoder (MHG-VAE), which uses a single VAE to achieve 100% validity. Our idea is to develop a graph grammar encoding the hard chemical constraints, called molecular hypergraph grammar (MHG), which guides VAE to always generate valid molecules. We also present an algorithm to construct MHG from a set of molecules.

Hidden units can play essential roles in modeling time-series having long-term dependency or on-linearity but make it difficult to learn associated parameters. Here we propose a way to learn such a time-series model by training a backward model for the time-reversed time-series, where the backward model has a common set of parameters as the original (forward) model. Our key observation is that only a subset of the parameters is hard to learn, and that subset is complementary between the forward model and the backward model. By training both of the two models, we can effectively learn the values of the parameters that are hard to learn if only either of the two models is trained. We apply bidirectional learning to a dynamic Boltzmann machine extended with hidden units. Numerical experiments with synthetic and real datasets clearly demonstrate advantages of bidirectional learning.

Dynamic Boltzmann machines (DyBMs) are recently developed generative models of a time series. They are designed to learn a time series by efficient online learning algorithms, whilst taking long-term dependencies into account with help of eligibility traces, recursively updatable memory units storing descriptive statistics of all the past data. The current DyBMs assume a finite-dimensional time series and cannot be applied to a functional time series, in which the dimension goes to infinity (e.g., spatiotemporal data on a continuous space). In this paper, we present a functional dynamic Boltzmann machine (F-DyBM) as a generative model of a functional time series. A technical challenge is to devise an online learning algorithm with which F-DyBM, consisting of functions and integrals, can learn a functional time series using only finite observations of it. We rise to the above challenge by combining a kernel-based function approximation method along with a statistical interpolation method and finally derive closed-form update rules. We design numerical experiments to empirically confirm the effectiveness of our solutions. The experimental results demonstrate consistent error reductions as compared to baseline methods, from which we conclude the effectiveness of F-DyBM for functional time series prediction.

Link prediction suffers from the data sparsity problem. This paper presents and validates our hypothesis that, for sparse networks, incidence matrix factorization (IMF) could perform better than adjacency matrix factorization (AMF), which has been used in many previous studies. A key observation supporting the hypothesis is that IMF models a partially-observed graph more accurately than AMF. A technical challenge for validating our hypothesis is that, unlike AMF approach, there does not exist an obvious method to make predictions using a factorized incidence matrix. To this end, we newly develop an optimization-based link prediction method adopting IMF. We have conducted thorough experiments using synthetic and real-world datasets to investigate the relationship between the sparsity of a network and the performance of the aforementioned two methods. The experimental results show that IMF performs better than AMF as networks become sparser, which strongly validates our hypothesis.

Knowledge on the Web relies heavily on multi-relational representations, such as RDF and Schema.org. Automatically extracting knowledge from documents and linking existing databases are common approaches to construct multi-relational data. Complementary to such approaches, there is still a strong demand for manually encoding human expert knowledge. For example, human annotation is necessary for constructing a common-sense knowledge base, which stores facts implicitly shared in a community, because such knowledge rarely appears in documents. As human annotation is both tedious and costly, an important research challenge is how to best use limited human resources, whiles maximizing the quality of the resulting dataset. In this paper, we formalize the problem of dataset construction as active learning problems and present the Active Multi-relational Data Construction (AMDC) method. AMDC repeatedly interleaves multi-relational learning and expert input acquisition, allowing us to acquire helpful labels for data construction. Experiments on real datasets demonstrate that our solution increases the number of positive triples by a factor of 2.28 to 17.0, and that the predictive performance of the multi-relational model in AMDC achieves the highest or comparable to the best performance throughout the data construction process.

Quality control is one of the central issues in crowdsourcing research. In this paper, we consider a quality control problem of crowdsourced enumeration tasks that request workers to enumerate possible answers as many as possible. Since workers neither necessarily provide correct answers nor provide exactly the same answers even if the answers indicate the same idea, we propose a two-stage quality control method consisting of the answer clustering stage and the reliability estimation stage.

Crowdsourcing is a technique to outsource tasks to a number of workers. Although crowdsourcing has many advantages, it gives rise to the risk that sensitive information may be leaked, which has limited the spread of its popularity. Task instances (data workers receive to process tasks) often contain sensitive information, which can be extracted by workers. For example, in an audio transcription task, an audio file corresponds to an instance, and the content of the audio (e.g., the abstract of a meeting) can be sensitive information. In this paper, we propose a quantitative analysis framework for the instance privacy problem. The proposed framework supplies us performance measures of instance privacy preserving protocols. As a case study, we apply the proposed framework to an instance clipping protocol and analyze the properties of the protocol. The protocol preserves privacy by clipping instances to limit the amount of information workers obtain. The results show that the protocol can balance task performance and instance privacy preservation. They also show that the proposed measure is consistent with standard measures, which validates the proposed measure.

We present a clustered personal classifier method (CPC method) that jointly estimates a classifier and clusters of workers in order to address the learning from crowds problem.Crowdsourcing allows us to create a large but low-quality data set at very low cost.The learning from crowds problem is to learn a classifier from such a low-quality data set.From some observations, we notice that workers form clusters according to their abilities.Although such a fact was pointed out several times, no method has applied it to the learning from crowds problem.We propose a CPC method that utilizes the clusters of the workers to improve the performance of the obtained classifier, where both the classifier and the clusters of the workers are estimated.The proposed method has two advantages.One is that it realizes robust estimation of the classifier because it utilizes prior knowledge about the workers that they tend to form clusters.The other is that we can obtain the clusters of the workers, which help us analyze the properties of the workers.Experimental results on synthetic and real data sets indicate that the proposed method can estimate the classifier robustly.In addition, clustering workers is shown to work well. Especially in the real data set, an outlier worker was found by applying the proposed method.

Recently crowdsourcing services are often used to collect a large amount of labeled data for machine learning, since they provide us an easy way to get labels at very low cost and in a short period. The use of crowdsourcing has introduced a new challenge in machine learning, that is, coping with the variable quality of crowd-generated data. Although there have been many recent attempts to address the quality problem of multiple workers, only a few of the existing methods consider the problem of learning classifiers directly from such noisy data. All these methods modeled the true labels as latent variables, which resulted in non-convex optimization problems. In this paper, we propose a convex optimization formulation for learning from crowds without estimating the true labels by introducing personal models of the individual crowd workers. We also devise an efficient iterative method for solving the convex optimization problems by exploiting conditional independence structures in multiple classifiers. We evaluate the proposed method against three competing methods on synthetic data sets and a real crowdsourced data set and demonstrate that the proposed method outperforms the other three methods.