Many real-world problems have complicated objective functions. To optimize such functions, humans utilize sophisticated sequential decision-making strategies. Many optimization algorithms have also been developed for this same purpose, but how do they compare to humans in terms of both performance and behavior? We try to unravel the general underlying algorithm people may be using while searching for the maximum of an invisible 1D function. Subjects click on a blank screen and are shown the ordinate of the function at each clicked abscissa location. Their task is to find the function’s maximum in as few clicks as possible. Subjects win if they get close enough to the maximum location. Analysis over 23 non-maths undergraduates, optimizing 25 functions from different families, shows that humans outperform 24 well-known optimization algorithms. Bayesian Optimization based on Gaussian Processes, which exploit all the x values tried and all the f(x) values obtained so far to pick the next x, predicts human performance and searched locations better. In 6 follow-up controlled experiments over 76 subjects, covering interpolation, extrapolation, and optimization tasks, we further confirm that Gaussian Processes provide a general and unified theoretical account to explain passive and active function learning and search in humans.

Psychophysical experiments have demonstrated that the brain integrates information from multiple sensory cues in a near Bayesian optimal manner. The present study proposes a novel mechanism to achieve this. We consider two reciprocally connected networks, mimicking the integration of heading direction information between the dorsal medial superior temporal (MSTd) and the ventral intraparietal (VIP) areas. Each network serves as a local estimator and receives an independent cue, either the visual or the vestibular, as direct input for the external stimulus. We find that positive reciprocal interactions can improve the decoding accuracy of each individual network as if it implements Bayesian inference from two cues. Our model successfully explains the experimental finding that both MSTd and VIP achieve Bayesian multisensory integration, though each of them only receives a single cue as direct external input. Our result suggests that the brain may implement optimal information integration distributively at each local estimator through the reciprocal connections between cortical regions.

Multi-step-ahead time series prediction is one of the most challenging research topics in the field of time series modeling and prediction, and is continually under research. Recently, the multiple-input several multiple-outputs (MISMO) modeling strategy has been proposed as a promising alternative for multi-step-ahead time series prediction, exhibiting advantages compared with the two currently dominating strategies, the iterated and the direct strategies. Built on the established MISMO strategy, this study proposes a particle swarm optimization (PSO)-based MISMO modeling strategy, which is capable of determining the number of sub-models in a self-adaptive mode, with varying prediction horizons. Rather than deriving crisp divides with equal-size s prediction horizons from the established MISMO, the proposed PSO-MISMO strategy, implemented with neural networks, employs a heuristic to create flexible divides with varying sizes of prediction horizons and to generate corresponding sub-models, providing considerable flexibility in model construction, which has been validated with simulated and real datasets.

Variational inference has become a widely used method to approximate posteriors in complex latent variables models. However, deriving a variational inference algorithm generally requires significant model-specific analysis, and these efforts can hinder and deter us from quickly developing and exploring a variety of models for a problem at hand. In this paper, we present a "black box" variational inference algorithm, one that can be quickly applied to many models with little additional derivation. Our method is based on a stochastic optimization of the variational objective where the noisy gradient is computed from Monte Carlo samples from the variational distribution. We develop a number of methods to reduce the variance of the gradient, always maintaining the criterion that we want to avoid difficult model-based derivations. We evaluate our method against the corresponding black box sampling based methods. We find that our method reaches better predictive likelihoods much faster than sampling methods. Finally, we demonstrate that Black Box Variational Inference lets us easily explore a wide space of models by quickly constructing and evaluating several models of longitudinal healthcare data.

In distributed ML applications, shared parameters are usually replicated among computing nodes to minimize network overhead. Therefore, proper consistency model must be carefully chosen to ensure algorithm's correctness and provide high throughput. Existing consistency models used in general-purpose databases and modern distributed ML systems are either too loose to guarantee correctness of the ML algorithms or too strict and thus fail to fully exploit the computing power of the underlying distributed system. Many ML algorithms fall into the category of \emph{iterative convergent algorithms} which start from a randomly chosen initial point and converge to optima by repeating iteratively a set of procedures. We've found that many such algorithms are to a bounded amount of inconsistency and still converge correctly. This property allows distributed ML to relax strict consistency models to improve system performance while theoretically guarantees algorithmic correctness. In this paper, we present several relaxed consistency models for asynchronous parallel computation and theoretically prove their algorithmic correctness. The proposed consistency models are implemented in a distributed parameter server and evaluated in the context of a popular ML application: topic modeling.

Consider observing an undirected network that is `noisy' in the sense that there are Type I and Type II errors in the observation of edges. Such errors can arise, for example, in the context of inferring gene regulatory networks in genomics or functional connectivity networks in neuroscience. Given a single observed network then, to what extent are summary statistics for that network representative of their analogues for the true underlying network? Can we infer such statistics more accurately by taking into account the noise in the observed network edges? In this paper, we answer both of these questions. In particular, we develop a spectral-based methodology using the adjacency matrix to `denoise' the observed network data and produce more accurate inference of the summary statistics of the true network. We characterize performance of our methodology through bounds on appropriate notions of risk in the $L^2$ sense, and conclude by illustrating the practical impact of this work on synthetic and real-world data.

Gaussian processes are rich distributions over functions, which provide a Bayesian nonparametric approach to smoothing and interpolation. We introduce simple closed form kernels that can be used with Gaussian processes to discover patterns and enable extrapolation. These kernels are derived by modelling a spectral density -- the Fourier transform of a kernel -- with a Gaussian mixture. The proposed kernels support a broad class of stationary covariances, but Gaussian process inference remains simple and analytic. We demonstrate the proposed kernels by discovering patterns and performing long range extrapolation on synthetic examples, as well as atmospheric CO2 trends and airline passenger data. We also show that we can reconstruct standard covariances within our framework.

This paper proposes the formalization and implementation of a novel class of constraints aimed at modeling problems related to placement of multi-body systems in the 3-dimensional space. Each multi-body is a system composed of body elements, connected by joint relationships and constrained by geometric properties. The emphasis of this investigation is the use of multi-body systems to model native conformations of protein structures---where each body represents an entity of the protein (e.g., an amino acid, a small peptide) and the geometric constraints are related to the spatial properties of the composing atoms. The paper explores the use of the proposed class of constraints to support a variety of different structural analysis of proteins, such as loop modeling and structure prediction. The declarative nature of a constraint-based encoding provides elaboration tolerance and the ability to make use of any additional knowledge in the analysis studies. The filtering capabilities of the proposed constraints also allow to control the number of representative solutions that are withdrawn from the conformational space of the protein, by means of criteria driven by uniform distribution sampling principles. In this scenario it is possible to select the desired degree of precision and/or number of solutions. The filtering component automatically excludes configurations that violate the spatial and geometric properties of the composing multi-body system. The paper illustrates the implementation of a constraint solver based on the multi-body perspective and its empirical evaluation on protein structure analysis problems.

Multi-task learning has shown to significantly enhance the performance of multiple related learning tasks in a variety of situations. We present the fused logistic regression, a sparse multi-task learning approach for binary classification. Specifically, we introduce sparsity inducing penalties over parameter differences of related logistic regression models to encode similarity across related tasks. The resulting joint learning task is cast into a form that lends itself to be efficiently optimized with a recursive variant of the alternating direction method of multipliers. We show results on synthetic data and describe the regime of settings where our multi-task approach achieves significant improvements over the single task learning approach and discuss the implications on applying the fused logistic regression in different real world settings.

Training large machine learning (ML) models with many variables or parameters can take a long time if one employs sequential procedures even with stochastic updates. A natural solution is to turn to distributed computing on a cluster; however, naive, unstructured parallelization of ML algorithms does not usually lead to a proportional speedup and can even result in divergence, because dependencies between model elements can attenuate the computational gains from parallelization and compromise correctness of inference. Recent efforts toward this issue have benefited from exploiting the static, a priori block structures residing in ML algorithms. In this paper, we take this path further by exploring the dynamic block structures and workloads therein present during ML program execution, which offers new opportunities for improving convergence, correctness, and load balancing in distributed ML. We propose and showcase a general-purpose scheduler, STRADS, for coordinating distributed updates in ML algorithms, which harnesses the aforementioned opportunities in a 1 systematic way. We provide theoretical guarantees for our scheduler, and demonstrate its efficacy versus static block structures on Lasso and Matrix Factorization. 1. INTRODUCTION Sensory techniques and digital storage media have improved at a breakneck pace, leading to massive collections of data. The resultant so-called Big Data problems have been a common focus in recent enthusiasms toward scalable machine learning, and numerous algorithmic and system solutions have been proposed to alleviate the time-bottleneck due to Big Data by exploring various heuristic or principled strategies for data parallelism [3, 18, 20, 28].