Optimal Transport (OT) distances are now routinely used as loss functions in ML tasks. Yet, computing OT distances between arbitrary (i.e. not necessarily discrete) probability distributions remains an open problem. This paper introduces a new online estimator of entropy-regularized OT distances between two such arbitrary distributions. It uses streams of samples from both distributions to iteratively enrich a non-parametric representation of the transportation plan. Compared to the classic Sinkhorn algorithm, our method leverages new samples at each iteration, which enables a consistent estimation of the true regularized OT distance. We cast our algorithm as a block-convex mirror descent in the space of positive distributions, and provide a theoretical analysis of its convergence. We numerically illustrate the performance of our method in comparison with concurrent approaches.

Deep reinforcement learning (RL) algorithms have recently achieved remarkable successes in various sequential decision making tasks, leveraging advances in methods for training large deep networks. However, these methods usually require large amounts of training data, which is often a big problem for real-world applications. One natural question to ask is whether learning good representations for states and using larger networks helps in learning better policies. In this paper, we try to study if increasing input dimensionality helps improve performance and sample efficiency of model-free deep RL algorithms. To do so, we propose an online feature extractor network (OFENet) that uses neural nets to produce good representations to be used as inputs to deep RL algorithms. Even though the high dimensionality of input is usually supposed to make learning of RL agents more difficult, we show that the RL agents in fact learn more efficiently with the high-dimensional representation than with the lower-dimensional state observations. We believe that stronger feature propagation together with larger networks (and thus larger search space) allows RL agents to learn more complex functions of states and thus improves the sample efficiency. Through numerical experiments, we show that the proposed method outperforms several other state-of-the-art algorithms in terms of both sample efficiency and performance.

To effectively train Takagi-Sugeno-Kang (TSK) fuzzy systems for regression problems, a Mini-Batch Gradient Descent with Regularization, DropRule, and AdaBound (MBGD-RDA) algorithm was recently proposed. It has demonstrated superior performances; however, there are also some limitations, e.g., it does not allow the user to specify the number of rules directly, and only Gaussian MFs can be used. This paper proposes two variants of MBGD-RDA to remedy these limitations, and show that they outperform the original MBGD-RDA and the classical ANFIS algorithms with the same number of rules. Furthermore, we also propose a rule pruning algorithm for TSK fuzzy systems, which can reduce the number of rules without significantly sacrificing the regression performance. Experiments showed that the rules obtained from pruning are generally better than training them from scratch directly, especially when Gaussian MFs are used.

This paper proposes a novel variational autoencoder (VAE) derived from Tsallis statistics, named q-VAE. A vanilla VAE is utilized to statistically extract latent space hidden in data sampled. Such latent space is useful to make robots controllable in feasible computational time and cost. To improve usefulness of the latent space, this paper focuses on disentangled representation learning like $\beta$-VAE, which is the baseline for it. Starting from the viewpoint of Tsallis statistics, a new lower bound of the q-VAE is derived to maximize likelihood of the data sampled. This can be regarded as an adaptive $\beta$-VAE with a deformed Kullback-Leibler divergence. To verify benefits from the q-VAE, a benchmark task to extract the latent space from MNIST dataset is performed. It is found that the q-VAE improved the disentangled representation while not deteriorating reconstruction accuracy of the data. As another advantage of the q-VAE, it does not require independency between the data. This advantage is demonstrated in learning latent dynamics of a nonlinear dynamical simulation. By combining the disentangled representation, the q-VAE achieves stable and accurate long-term state prediction from the initial state and the actions at respective times.

Batch-normalization (BN) is a key component to effectively train deep neural networks. Empirical evidence has shown that without BN, the training process is prone to unstabilities. This is however not well understood from a theoretical point of view. Leveraging tools from Markov chain theory, we show that BN has a direct effect on the rank of the pre-activation matrices of a neural network. Specifically, while deep networks without BN exhibit rank collapse and poor training performance, networks equipped with BN have a higher rank. In an extensive set of experiments on standard neural network architectures and datasets, we show that the latter quantity is a good predictor for the optimization speed of training.

We consider bounded width CNF-formulas where the width is measured by popular graph width measures on graphs associated to CNF-formulas. Such restricted graph classes, in particular those of bounded treewidth, have been extensively studied for their uses in the design of algorithms for various computational problems on CNF-formulas. Here we consider the expressivity of these formulas in the model of clausal encodings with auxiliary variables. We first show that bounding the width for many of the measures from the literature leads to a dramatic loss of expressivity, restricting the formulas to those of low communication complexity. We then show that the width of optimal encodings with respect to different measures is strongly linked: there are two classes of width measures, one containing primal treewidth and the other incidence cliquewidth, such that in each class the width of optimal encodings only differs by constant factors. Moreover, between the two classes the width differs at most by a factor logarithmic in the number of variables. Both these results are in stark contrast to the setting without auxiliary variables where all width measures we consider here differ by more than constant factors and in many cases even by linear factors.

Crowdsourcing is a popular methodology to collect manual labels at scale. Such labels are often used to train AI models and, thus, quality control is a key aspect in the process. One of the most popular quality assurance mechanisms in paid micro-task crowdsourcing is based on gold questions: the use of a small set of tasks of which the requester knows the correct answer and, thus, is able to directly assess crowd work quality. In this paper, we show that such mechanism is prone to an attack carried out by a group of colluding crowd workers that is easy to implement and deploy: the inherent size limit of the gold set can be exploited by building an inferential system to detect which parts of the job are more likely to be gold questions. The described attack is robust to various forms of randomisation and programmatic generation of gold questions. We present the architecture of the proposed system, composed of a browser plug-in and an external server used to share information, and briefly introduce its potential evolution to a decentralised implementation. We implement and experimentally validate the gold detection system, using real-world data from a popular crowdsourcing platform.  Our experimental results show that crowdworkers using the proposed system spend more time on signalled gold questions but do not neglect the others thus achieving an increased overall work quality. Finally, we discuss the economic and sociological implications of this kind of attack.

Recent empirical works show that large deep neural networks are often highly redundant and one can find much smaller subnetworks without a significant drop of accuracy. However, most existing methods of network pruning are empirical and heuristic, leaving it open whether good subnetworks provably exist, how to find them efficiently, and if network pruning can be provably better than direct training using gradient descent. We answer these problems positively by proposing a simple greedy selection approach for finding good subnetworks, which starts from an empty network and greedily adds important neurons from the large network. This differs from the existing methods based on backward elimination, which remove redundant neurons from the large network. Theoretically, applying our greedy selection strategy on sufficiently large pre-trained networks guarantees to find small subnetworks with lower loss than networks directly trained with gradient descent. Practically, we improve prior arts of network pruning on learning compact neural architectures on ImageNet, including ResNet, MobilenetV2/V3, and ProxylessNet. Our theory and empirical results on MobileNet suggest that we should fine-tune the pruned subnetworks to leverage the information from the large model, instead of re-training from new random initialization as suggested in \citet{liu2018rethinking}.

Convolutional Neural Networks (CNNs) have shown impressive performance in computer vision tasks such as image classification and segmentation. One factor for the success of CNNs is that they have an inductive bias that assumes a certain type of spatial structure is present in the data. Recent work by Geirhos et al. (2018) shows how learning in CNNs causes the learned CNN models to be biased towards high-frequency textural information, compared to low-frequency shape information in images. Many tasks generally requires both shape and textural information. Hence, we propose a simple curriculum based scheme which improves the ability of CNNs to be less biased towards textural information, and at the same time, being able to represent both the shape and textural information. We propose to augment the training of CNNs by controlling the amount of textural information that is available to the CNNs during the training process, by convolving the output of a CNN layer with a low-pass filter, or simply a Gaussian kernel. By reducing the standard deviation of the Gaussian kernel, we are able to gradually increase the amount of textural information available as training progresses, and hence reduce the texture bias. Such an augmented training scheme significantly improves the performance of CNNs on various image classification tasks, while adding no additional trainable parameters or auxiliary regularization objectives. We also observe significant improvements when using the trained CNNs to perform transfer learning on a different dataset, and transferring to a different task which shows how the learned CNNs using the proposed method act as better feature extractors.

In recent decades, attention has been directed at anemia classification for various medical purposes, such as thalassemia screening and predicting iron deficiency anemia (IDA). In this study, a new method has been successfully tested for discrimination between IDA and \b{eta}-thalassemia trait (\b{eta}-TT). The method is based on a Dynamic Harmony Search (DHS). Complete blood count (CBC), a fast and inexpensive laboratory test, is used as the input of the system. Other models, such as a genetic programming method called structured representation on genetic algorithm in non-linear function fitting (STROGANOFF), an artificial neural network (ANN), an adaptive neuro-fuzzy inference system (ANFIS), a support vector machine (SVM), k-nearest neighbor (KNN), and certain traditional methods, are compared with the proposed method.