Manifold hypotheses are typically used for tasks such as dimensionality reduction, interpolation, or improving classification performance. In the less common problem of manifold estimation, the task is to characterize the geometric structure of the manifold in the original ambient space from a sample. We focus on the role that tangent bundle learners (TBL) can play in estimating the underlying manifold from which data is assumed to be sampled. Since the unbounded tangent spaces natively represent a poor manifold estimate, the problem reduces to one of estimating regions in the tangent space where it acts as a relatively faithful linear approximator to the surface of the manifold. Local PCA methods, such as the Mixtures of Probabilistic Principal Component Analyzers method of Tipping and Bishop produce a subset of the tangent bundle of the manifold along with an assignment function that assigns points in the training data used by the TBL to elements of the estimated tangent bundle. We formulate three methods that use the data assigned to each tangent space to estimate the underlying bounded subspaces for which the tangent space is a faithful estimate of the manifold and offer thoughts on how this perspective is theoretically grounded in the manifold assumption. We seek to explore the conceptual and technical challenges that arise in trying to utilize simple TBL methods to arrive at reliable estimates of the underlying manifold.

A wealth of algorithms have been developed to extract natural cluster structure in data. Identifying this structure is desirable but not always sufficient: We may also want to understand why the data points have been assigned to a given cluster. Clustering algorithms do not offer a systematic answer to this simple question. Hence we propose a new framework that can, for the first time, explain cluster assignments in terms of input features in a comprehensive manner. It is based on the novel theoretical insight that clustering models can be rewritten as neural networks, or 'neuralized'. Predictions of the obtained networks can then be quickly and accurately attributed to the input features. Several showcases demonstrate the ability of our method to assess the quality of learned clusters and to extract novel insights from the analyzed data and representations.

In neural architecture search, the structure of the neural network to best model a given dataset is determined by an automated search process. Efficient Neural Architecture Search (ENAS), proposed by Pham et al. (2018), has recently received considerable attention due to its ability to find excellent architectures within a comparably short search time. In this work, which is motivated by the quest to further improve the learning speed of architecture search, we evaluate the learning progress of the controller which generates the architectures in ENAS. We measure the progress by comparing the architectures generated by it at different controller training epochs, where architectures are evaluated after having re-trained them from scratch. As a surprising result, we find that the learning curves are completely flat, i.e., there is no observable progress of the controller in terms of the performance of its generated architectures. This observation is consistent across the CIFAR-10 and CIFAR-100 datasets and two different search spaces. We conclude that the high quality of the models generated by ENAS is a result of the search space design rather than the controller training, and our results indicate that one-shot architecture design is an efficient alternative to architecture search by ENAS.

Analyzing mobility behavior of users is extremely useful to create or improve existing services. Several research works have been done in order to study mobility behavior of users that mainly use users' significant locations. However, these existing analysis are extremely intrusive because they require the knowledge of the frequently visited places of users, which thus makes it fairly easy to identify them. Consequently, in this paper, we present a privacy-aware methodology to analyze mobility behavior of users. We firstly propose a new metric based on the well-known Shannon entropy, called spatio-temporal entropy, to quantify the mobility level of a user during a time window. Then, we compute a sequence of spatio-temporal entropy from the location history of the user that expresses user's movements as rhythms. We secondly present how to study the effects of several groups of additional variables on the evolution of the spatio-temporal entropy of a user, such as spatio-temporal, demographic and mean of transportation variables. For this, we use Generalized Additive Models (GAMs). The results firstly show that the spatio-temporal entropy and GAMs are an ideal combination to understand mobility behavior of an individual user or a group of users. We also evaluate the prediction accuracy of a global GAM compared to individual GAMs and individual AutoRegressive Integrated Moving Average (ARIMA) models. These last results highlighted that the global GAM gives more accurate predictions of spatio-temporal entropy by checking the Mean Absolute Error (MAE). In addition, this research work opens various threads, such as the prediction of demographic data of users or the creation of personalized mobility prediction models by using movement rhythm characteristics of a user.

While recent NAS algorithms are thousands of times faster than the pioneering works, it is often overlooked that they use fewer candidate operations, resulting in a significantly smaller search space. We present PR-DARTS, a NAS algorithm that discovers strong network configurations in a much larger search space and a single day. A small candidate operation pool is used, from which candidates are progressively pruned and replaced with better performing ones. Experiments on CIFAR-10 and CIFAR-100 achieve 2.51% and 15.53% test error, respectively, despite searching in a space where each cell has 150 times as many possible configurations than in the DARTS baseline.

With the explosive growth of e-commerce and the booming of e-payment, detecting online transaction fraud in real time has become increasingly important to Fintech business. To tackle this problem, we introduce the TitAnt, a transaction fraud detection system deployed in Ant Financial, one of the largest Fintech companies in the world. The system is able to predict online real-time transaction fraud in mere milliseconds. We present the problem definition, feature extraction, detection methods, implementation and deployment of the system, as well as empirical effectiveness. Extensive experiments have been conducted on large real-world transaction data to show the effectiveness and the efficiency of the proposed system.

Imitation from observation is the framework of learning tasks by observing demonstrated state-only trajectories. Recently, adversarial approaches have achieved significant performance improvements over other methods for imitating complex behaviors. However, these adversarial imitation algorithms often require many demonstration examples and learning iterations to produce a policy that is successful at imitating a demonstrator's behavior. This high sample complexity often prohibits these algorithms from being deployed on physical robots. In this paper, we propose an algorithm that addresses the sample inefficiency problem by utilizing ideas from trajectory centric reinforcement learning algorithms. We test our algorithm and conduct experiments using an imitation task on a physical robot arm and its simulated version in Gazebo and will show the improvement in learning rate and efficiency.

Imitation learning has long been an approach to alleviate the tractability issues that arise in reinforcement learning. However, most literature makes several assumptions such as access to the expert's actions, availability of many expert demonstrations, and injection of task-specific domain knowledge into the learning process. We propose reinforced inverse dynamics modeling (RIDM), a method of combining reinforcement learning and imitation from observation (IfO) to perform imitation using a single expert demonstration, with no access to the expert's actions, and with little task-specific domain knowledge. Given only a single set of the expert's raw states, such as joint angles in a robot control task, at each time-step, we learn an inverse dynamics model to produce the necessary low-level actions, such as torques, to transition from one state to the next such that the reward from the environment is maximized. We demonstrate that RIDM outperforms other techniques when we apply the same constraints on the other methods on six domains of the MuJoCo simulator and for two different robot soccer tasks for two experts from the RoboCup 3D simulation league on the SimSpark simulator.

We introduce a simple framework for identifying biases of a smiling attribute classifier. Our method poses counterfactual questions of the form: how would the prediction change if this face characteristic had been different? We leverage recent advances in generative adversarial networks to build a realistic generative model of face images that affords controlled manipulation of specific image characteristics. We introduce a set of metrics that measure the effect of manipulating a specific property of an image on the output of a trained classifier. Empirically, we identify several different factors of variation that affect the predictions of a smiling classifier trained on CelebA.

Recurrent neural networks can be large and compute-intensive, yet many applications that benefit from RNNs run on small devices with very limited compute and storage capabilities while still having run-time constraints. As a result, there is a need for compression techniques that can achieve significant compression without negatively impacting inference run-time and task accuracy. This paper explores a new compressed RNN cell implementation called Hybrid Matrix Decomposition (HMD) that achieves this dual objective. This scheme divides the weight matrix into two parts - an unconstrained upper half and a lower half composed of rank-1 blocks. This results in output features where the upper sub-vector has "richer" features while the lower-sub vector has "constrained" features". HMD can compress RNNs by a factor of 2-4x while having a faster run-time than pruning and retaining more model accuracy than matrix factorization. We evaluate this technique on 3 benchmarks.