The scalable inference for Bayesian nonparametric models with big data is still challenging. Current variational inference methods fail to characterise the correlation structure among latent variables due to the mean-field setting and cannot infer the true posterior dimension because of the universal truncation. To overcome these limitations, we build a general framework to infer Bayesian nonparametric models by maximising the proposed nonparametric evidence lower bound, and then develop a novel approach by combining Monte Carlo sampling and stochastic variational inference framework. Our method has several advantages over the traditional online variational inference method. First, it achieves a smaller divergence between variational distributions and the true posterior by factorising variational distributions under the conditional setting instead of the mean-field setting to capture the correlation pattern. Second, it reduces the risk of underfitting or overfitting by truncating the dimension adaptively rather than using a prespecified truncated dimension for all latent variables. Third, it reduces the computational complexity by approximating the posterior functionally instead of updating the stick-breaking parameters individually. We apply the proposed method on hierarchical Dirichlet process and gamma--Dirichlet process models, two essential Bayesian nonparametric models in topic analysis. The empirical study on three large datasets including arXiv, New York Times and Wikipedia reveals that our proposed method substantially outperforms its competitor in terms of lower perplexity and much clearer topic-words clustering.

How does a 110-layer ResNet learn a high-complexity classifier using relatively few training examples and short training time? We present a theory towards explaining this in terms of $\textit{hierarchical learning}$. We refer hierarchical learning as the learner learns to represent a complicated target function by decomposing it into a sequence of simpler functions to reduce sample and time complexity. This paper formally analyzes how multi-layer neural networks can perform such hierarchical learning efficiently and automatically simply by applying stochastic gradient descent (SGD). On the conceptual side, we present, to the best of our knowledge, the FIRST theory result indicating how very deep neural networks can still be sample and time efficient on certain hierarchical learning tasks, when NO KNOWN non-hierarchical algorithms (such as kernel method, linear regression over feature mappings, tensor decomposition, sparse coding) are efficient. We establish a new principle called "backward feature correction", which we believe is the key to understand the hierarchical learning in multi-layer neural networks. On the technical side, we show for regression and even for binary classification, for every input dimension $d > 0$, there is a concept class consisting of degree $\omega(1)$ multi-variate polynomials so that, using $\omega(1)$-layer neural networks as learners, SGD can learn any target function from this class in $\mathsf{poly}(d)$ time using $\mathsf{poly}(d)$ samples to any $\frac{1}{\mathsf{poly}(d)}$ error, through learning to represent it as a composition of $\omega(1)$ layers of quadratic functions. In contrast, we present lower bounds stating that several non-hierarchical learners, including any kernel methods, neural tangent kernels, must suffer from $d^{\omega(1)}$ sample or time complexity to learn functions in this concept class even to any $d^{-0.01}$ error.

When people learn mathematical patterns or sequences, they are able to identify the concepts (or rules) underlying those patterns. Having learned the underlying concepts, humans are also able to generalize those concepts to other numbers, so far as to even identify previously unseen combinations of those rules. Current state-of-the art RNN architectures like LSTMs perform well in predicting successive elements of sequential data, but require vast amounts of training examples. Even with extensive data, these models struggle to generalize concepts. From our behavioral study, we also found that humans are able to disregard noise and identify the underlying rules generating the corrupted sequences. We therefore propose a Bayesian model that captures these human-like learning capabilities to predict next number in a given sequence, better than traditional LSTMs.

Estimating temporal patterns in travel times along road segments in urban settings is of central importance to traffic engineers and city planners. In this work, we propose a methodology to leverage coarse-grained and aggregated travel time data to estimate the street-level travel times of a given metropolitan area. Our main focus is to estimate travel times along the arterial road segments where relevant data are often unavailable. The central idea of our approach is to leverage easy-to-obtain, aggregated data sets with broad spatial coverage, such as the data published by Uber Movement, as the fabric over which other expensive, fine-grained datasets, such as loop counter and probe data, can be overlaid. Our proposed methodology uses a graph representation of the road network and combines several techniques such as graph-based routing, trip sampling, graph sparsification, and least-squares optimization to estimate the street-level travel times. Using sampled trips and weighted shortest-path routing, we iteratively solve constrained least-squares problems to obtain the travel time estimates. We demonstrate our method on the Los Angeles metropolitan-area street network, where aggregated travel time data is available for trips between traffic analysis zones. Additionally, we present techniques to scale our approach via a novel graph pseudo-sparsification technique.

Recent research in behaviour understanding through language grounding has shown it is possible to automatically generate behaviour models from textual instructions. These models usually have goal-oriented structure and are modelled with different formalisms from the planning domain such as the Planning Domain Definition Language. One major problem that still remains is that there are no benchmark datasets for comparing the different model generation approaches, as each approach is usually evaluated on domain-specific application. To allow the objective comparison of different methods for model generation from textual instructions, in this report we introduce a dataset consisting of 83 textual instructions in English language, their refinement in a more structured form as well as manually developed plans for each of the instructions. The dataset is publicly available to the community.

Fusion is a common tool for the analysis and utilization of available datasets and so an essential part of data mining and machine learning processes. However, a clear definition of the type of fusion is not always provided due to inconsistent literature. In the following, the process of fusion is defined depending on the fusion components and the abstraction level on which the fusion occurs. The focus in the first part of the paper at hand is on the clear definition of the terminology and the development of an appropriate ontology of the fusion components and the fusion level. In the second part, common fusion techniques are presented.

H. Kargupta: Connected Cars: How Distributed Data Mining Is Changing the Next Generation of Vehicle Telematics Products.

There is an increasing desire for autonomous systems to have high levels of robustness and safety, attained through continuously planning and self-repairing online. Underlying this is the need to accurately estimate the system state and diagnose subtle failures. Estimation methods based on hybrid discrete and continuous state models have emerged as a method of precisely computing these estimates. However, existing methods have difficulty scaling to systems with more than a handful of components. Discrete, consistency based state estimation capabilities can scale to this level by combining best-first enumeration and conflict-directed search. While best-first methods have been developed for hybrid estimation, conflict-directed methods have thus far been elusive as conflicts learn inconsistencies from constraint violation, but probabilistic hybrid estimation is relatively unconstrained. In this paper we present an approach to hybrid estimation that unifies best-first enumeration and conflict-directed search through the concept of "bounding" conflicts, an extension of conflicts that represent tighter bounds on the cost of regions of the search space. This paper presents a general best-first enumeration algorithm based on bounding conflicts (A*BC) and a hybrid estimation method using this enumeration algorithm. Experiments show that an A*BC powered state estimator produces estimates up to an order of magnitude faster than the current state of the art, particularly on large systems.

Many medical decision-making settings can be framed as partially observed Markov decision processes (POMDPs). However, popular two-stage approaches that first learn a POMDP model and then solve it often fail because the model that best fits the data may not be the best model for planning. We introduce a new optimization objective that (a) produces both high-performing policies and high-quality generative models, even when some observations are irrelevant for planning, and (b) does so in the kinds of batch, off-policy settings common in medicine. We demonstrate our approach on synthetic examples and a real-world hypotension management task.

Heterogeneity between training and testing data degrades reproducibility of a well-trained predictive algorithm. In modern applications, how to deploy a trained algorithm in a different domain is becoming an urgent question raised by many domain scientists. In this paper, we propose a reproducible bootstrap aggregating (Rbagging) method coupled with a new algorithm, the iterative nearest neighbor sampler (INNs), effectively drawing bootstrap samples from training data to mimic the distribution of the test data. Rbagging is a general ensemble framework that can be applied to most classifiers. We further propose Rbagging+ to effectively detect anomalous samples in the testing data. Our theoretical results show that the resamples based on Rbagging have the same distribution as the testing data. Moreover, under suitable assumptions, we further provide a general bound to control the test excess risk of the ensemble classifiers. The proposed method is compared with several other popular domain adaptation methods via extensive simulation studies and real applications including medical diagnosis and imaging classifications.