To guide their behavior, humans and animals rely on previously learned knowledge about the world. Since the world is complex and models of the world are never perfect, the question arises whether we should trust our internal world model that we have built from past data or whether we should readjust it when we receive a new data sample. In noisy environments, a single data sample may not be reliable and in general we need to average over several data samples. However, when a structural change occurs in the environment, the most recent data samples are the most informative ones and we should put more weight on recent data samples than on earlier ones. Indeed, both humans and animals adaptively adjust the relative contribution of old and newly acquired data during learning (Behrens et al., 2007; Nassar et al., 2012; Krugel et al., 2009; Pearce and Hall, 1980) and rapidly adapt to changing environments (Pearce and Hall, 1980; Wilson et al., 1992; Holland, 1997).

The Allstate Claims Severity recruiting competition ran on Kaggle from October to December 2016. As Kaggle's most popular recruiting competitions to-date, it attracted over 3,000 entrants who competed to predict the loss value associated with Allstate insurance claims. In this interview, Alexey Noskov walks us through how he came in second place by creating features based on distance from cluster centroids and applying newfound intuitions for (hyper)-parameter tuning. Along the way, he provides details on his favorite tips and tricks including lots of feature engineering and implementing a custom objective function for XGBoost. I have MSc in computer science and work as a software engineer at Evil Martians.

Recently, technologies such as face detection, facial landmark localisation and face recognition and verification have matured enough to provide effective and efficient solutions for imagery captured under arbitrary conditions (referred to as "in-the-wild"). This is partially attributed to the fact that comprehensive "in-the-wild" benchmarks have been developed for face detection, landmark localisation and recognition/verification. A very important technology that has not been thoroughly evaluated yet is deformable face tracking "in-the-wild". Until now, the performance has mainly been assessed qualitatively by visually assessing the result of a deformable face tracking technology on short videos. In this paper, we perform the first, to the best of our knowledge, thorough evaluation of state-of-the-art deformable face tracking pipelines using the recently introduced 300VW benchmark. We evaluate many different architectures focusing mainly on the task of on-line deformable face tracking. In particular, we compare the following general strategies: (a) generic face detection plus generic facial landmark localisation, (b) generic model free tracking plus generic facial landmark localisation, as well as (c) hybrid approaches using state-of-the-art face detection, model free tracking and facial landmark localisation technologies. Our evaluation reveals future avenues for further research on the topic.

We present an asymptotic criterion to determine the optimal number of clusters in k-means. We consider k-means as data compression, and propose to adopt the number of clusters that minimizes the estimated description length after compression. Here we report two types of compression ratio based on two ways to quantify the description length of data after compression. This approach further offers a way to evaluate whether clusters obtained with k-means have a hierarchical structure by examining whether multi-stage compression can further reduce the description length. We applied our criteria to determine the number of clusters to synthetic data and empirical neuroimaging data to observe the behavior of the criteria across different types of data set and suitability of the two types of criteria for different datasets. We found that our method can offer reasonable clustering results that are useful for dimension reduction. While our numerical results revealed dependency of our criteria on the various aspects of dataset such as the dimensionality, the description length approach proposed here provides a useful guidance to determine the number of clusters in a principled manner when underlying properties of the data are unknown and only inferred from observation of data.

Conducting surveillance missions using sensor networks is essential for many civilian and military applications, such as disaster response [1], border patrol [2], force protection [3], [4], combat missions [5], and traffic management [6]. One main task in these missions is to collect data regarding the operational environment and then obtain intelligence information from the data. Because the sensors used to collect data are often spatially distributed, extracting data patterns becomes critical to obtain accurate knowledge about the underlying activities. The existing work on identifying data patterns from spatially distributed sensors is focused on developing probabilistic reasoning techniques without recognizing the specific data association or data patterns. Existing approaches for multitarget state estimation can be characterized by two features: a data-to-target assignment algorithm, and an algorithm for single target state estimation under preexisting data-to-target associations. With unknown data association, probabilistic data association (PDA) [7] and multiple hypothesis tracking (MHT) [8] are two common approaches where dense measurements are available. In the study of traffic patterns, the existing research is focused on estimating traffic density and smart routes [6] without analyzing the data pattern to obtain better knowledge of traffic information.

Discrimination-aware classification is receiving an increasing attention in data science fields. The preprocess methods for constructing a discriminationfree classifier first remove discrimination from the training data, and then learn the classifier from the cleaned data. However, they lack a theoretical guarantee for the potential discrimination when the classifier is deployed for prediction. In this paper, we fill this gap by mathematically bounding the probability of the discrimination in prediction being within a given interval in terms of the training data and classifier. We adopt the causal model for modeling the data generation mechanism, and formally defining discrimination in population, in a dataset, and in prediction. We obtain two important theoretical results: (1) the discrimination in prediction can still exist even if the discrimination in the training data is completely removed; and (2) not all pre-process methods can ensure nondiscrimination in prediction even though they can achieve non-discrimination in the modified training data. Based on the results, we develop a two-phase framework for constructing a discrimination-free classifier with a theoretical guarantee. The experiments demonstrate the theoretical results and show the effectiveness of our two-phase framework.

Supervised machine learning methods are algorithms for inductive inference. On the basis of a sample, they construct (learn) a computable model of a data generating process that facilitates inference over the underlying ground truth function and aims to predict its function values at unobserved inputs. Among supervised learning methods, nonparametric algorithms tend to offer greater flexibility to learn rich function classes. Unfortunately, many classical techniques for nonparametric regression, such as the Nadaraya-Watson estimator [21], [14] or the LOESS method, [6] suffer from a practical limitation: their regression performance depends on the choice of hyperparameters. While in principle, it would be possible to tune these to the data (in manner similar in spirit to the one we propose in this work), to the best of our knowledge, currently there is little understanding on how to do so with a global optimiser that offers theoretical performance guarantees on the optimisation solution. This means that in practice, one is left to engineer these hyperparameters (or the settings of an optimiser) by manual tuning in order to ensure good performance on a particular learning problem. Of course, this stands in opposition to the motivation for utilising nonparametric learning, especially in system identification: which is to facilitate flexible and fully automated black-box learning that does not require manual intervention.

We study mechanisms to characterize how the asymptotic convergence of backpropagation in deep architectures, in general, is related to the network structure, and how it may be influenced by other design choices including activation type, denoising and dropout rate. We seek to analyze whether network architecture and input data statistics may guide the choices of learning parameters and vice versa. Given the broad applicability of deep architectures, this issue is interesting both from theoretical and a practical standpoint. Using properties of general nonconvex objectives (with first-order information), we first build the association between structural, distributional and learnability aspects of the network vis-\`a-vis their interaction with parameter convergence rates. We identify a nice relationship between feature denoising and dropout, and construct families of networks that achieve the same level of convergence. We then derive a workflow that provides systematic guidance regarding the choice of network sizes and learning parameters often mediated4 by input statistics. Our technical results are corroborated by an extensive set of evaluations, presented in this paper as well as independent empirical observations reported by other groups. We also perform experiments showing the practical implications of our framework for choosing the best fully-connected design for a given problem.

We study the estimation of the latent variable Gaussian graphical model (LVGGM), where the precision matrix is the superposition of a sparse matrix and a low-rank matrix. In order to speed up the estimation of the sparse plus low-rank components, we propose a sparsity constrained maximum likelihood estimator based on matrix factorization, and an efficient alternating gradient descent algorithm with hard thresholding to solve it. Our algorithm is orders of magnitude faster than the convex relaxation based methods for LVGGM. In addition, we prove that our algorithm is guaranteed to linearly converge to the unknown sparse and low-rank components up to the optimal statistical precision. Experiments on both synthetic and genomic data demonstrate the superiority of our algorithm over the state-of-the-art algorithms and corroborate our theory.

Graphlets are induced subgraphs of a large network and are important for understanding and modeling complex networks. Despite their practical importance, graphlets have been severely limited to applications and domains with relatively small graphs. Most previous work has focused on exact algorithms, however, it is often too expensive to compute graphlets exactly in massive networks with billions of edges, and finding an approximate count is usually sufficient for many applications. In this work, we propose an unbiased graphlet estimation framework that is (a) fast with significant speedups compared to the state-of-the-art, (b) parallel with nearly linear-speedups, (c) accurate with <1% relative error, (d) scalable and space-efficient for massive networks with billions of edges, and (e) flexible for a variety of real-world settings, as well as estimating macro and micro-level graphlet statistics (e.g., counts) of both connected and disconnected graphlets. In addition, an adaptive approach is introduced that finds the smallest sample size required to obtain estimates within a given user-defined error bound. On 300 networks from 20 domains, we obtain <1% relative error for all graphlets. This is significantly more accurate than existing methods while using less data. Moreover, it takes a few seconds on billion edge graphs (as opposed to days/weeks). These are by far the largest graphlet computations to date.