In online advertisement as well as many other fields such as health informatics and computational finance, we often have to deal with the situation in which we are given a batch of data generated by the current strategy(ies) of the company (hospital, investor), and we are asked to generate a good or an optimal strategy. Although there are many techniques to find a good policy given a batch of data, there are not much results to guarantee that the obtained policy will perform well in the real system without deploying it. On the other hand, deploying a policy might be risky, and thus, requires convincing the product (hospital, investment) manager that it is not going to harm the business. This is why it is extremely important to devise algorithms that generate policies with performance guarantees.

In this talk, we discuss four different approaches to this fundamental problem, we call them model-based, model-free, online, and risk-sensitive. In the model-based approach, we first use the batch of data and build a simulator that mimics the behavior of the dynamical system under studies (online advertisement, hospital’s ER, financial market), and then use this simulator to generate data and learn a policy. The main challenge here is to have guarantees on the performance of the learned policy, given the error in the simulator. This line of research is closely related to the area of robust learning and control. In the model-free approach, we learn a policy directly from the batch of data (without building a simulator), and the main question is whether the learned policy is guaranteed to perform at least as well as a baseline strategy. This line of research is related to off-policy evaluation and control. In the online approach, the goal is to control the exploration of the algorithm in a way that never during its execution the loss of using it instead of the baseline strategy is more than a given margin. In the risk-sensitive approach, the goal is to learn a policy that manages risk by minimizing some measure of variability in the performance in addition to maximizing a standard criterion. We present algorithms based on these approaches and demonstrate their usefulness in real-world applications such as personalized ad recommendation, energy arbitrage, traffic signal control, and American option pricing.

Bio:Mohammad Ghavamzadeh received a Ph.D. degree in Computer Science from the University of Massachusetts Amherst in 2005. From 2005 to 2008, he was a postdoctoral fellow at the University of Alberta. He has been a permanent researcher at INRIA in France since November 2008. He was promoted to first-class researcher in 2010, was the recipient of the “INRIA award for scientific excellence” in 2011, and obtained his Habilitation in 2014. He is currently (from October 2013) on a leave of absence from INRIA working as a senior analytics researcher at Adobe Research in California, on projects related to digital marketing. He has been an area chair and a senior program committee member at NIPS, IJCAI, and AAAI. He has been on the editorial board of Machine Learning Journal (MLJ), has published over 50 refereed papers in major machine learning, AI, and control journals and conferences, and has organized several tutorials and workshops at NIPS, ICML, and AAAI. His research is mainly focused on sequential decision-making under uncertainty, reinforcement learning, and online learning.

In this talk, I will first introduce a broad class of unsupervised deep learning models and show that they can learn useful hierarchical representations from large volumes of high-dimensional data with applications in information retrieval, object recognition, and speech perception. I will next introduce deep models that are capable of extracting a unified representation that fuses together multiple data modalities and present the Reverse Annealed Importance Sampling Estimator (RAISE) for evaluating these deep generative models. Finally, I will discuss models that can generate natural language descriptions (captions) of images and generate images from captions using attention, as well as introduce multiplicative and fine-grained gating mechanisms with application to reading comprehension.

Bio: Ruslan Salakhutdinov received his PhD in computer science from the University of Toronto in 2009. After spending two post-doctoral years at the Massachusetts Institute of Technology Artificial Intelligence Lab, he joined the University of Toronto as an Assistant Professor in the Departments of Statistics and Computer Science. In 2016 he joined the Machine Learning Department at Carnegie Mellon University as an Associate Professor. Ruslan’s primary interests lie in deep learning, machine learning, and large-scale optimization. He is an action editor of the Journal of Machine Learning Research and served on the senior programme committee of several learning conferences including NIPS and ICML. He is an Alfred P. Sloan Research Fellow, Microsoft Research Faculty Fellow, Canada Research Chair in Statistical Machine Learning, a recipient of the Early Researcher Award, Google Faculty Award, Nvidia’s Pioneers of AI award, and is a Senior Fellow of the Canadian Institute for Advanced Research.

Learning in the Machine is a style of machine learning that takes into account the physical constraints of learning machines, from brains to neuromorphic chips. Taking into account these constraints leads to new insights into the foundations of learning systems, and occasionally leads also to improvements for machine learning performed on digital computers. Learning in the Machine is particularly useful when applied to message passing algorithms such as backpropagation and belief propagation, and leads to the concepts of local learning and learning channel. These concepts in turn will be applied to random backpropagation and several new variants. In addition to simulations corroborating the remarkable robustness of these algorithms, we will present new mathematical results establishing interesting connections between machine learning and Hilbert 16th problem.

Tensor networks are a technique for factorizing tensors with hundreds or thousands of indices into a contracted network of low-order tensors. Originally developed at UCI in the 1990’s, tensor networks have revolutionized major areas of physics are starting to be used in applied math and machine learning. I will show that tensor networks fit naturally into a certain class of non-linear kernel learning models, such that advanced optimization techniques from physics can be applied straightforwardly (arxiv:1605.05775). I will discuss many advantages and future directions of tensor network models, for example adaptive pruning of weights and linear scaling with training set size (compared to at least quadratic scaling when using the kernel trick).

Bounding the partition function is a key inference task in many graphical models. In this paper, we develop an anytime anyspace search algorithm taking advantage of AND/OR tree structure and optimized variational heuristics to tighten deterministic bounds on the partition function. We study how our priority-driven best-first search scheme can improve on state-of-the-art variational bounds in an anytime way within limited memory resources, as well as the effect of the AND/OR framework to exploit conditional independence structure within the search process within the context of summation. We compare our resulting bounds to a number of existing methods, and show that our approach offers a number of advantages on real-world problem instances taken from recent UAI competitions.

Deep generative models (such as the Variational Autoencoder) efficiently couple the expressiveness of deep neural networks with the robustness to uncertainty of probabilistic latent variables. This talk will first give an overview of deep generative models, their applications, and approximate inference strategies for them. Then I’ll discuss our work on placing Bayesian Nonparametric priors on their latent space, which allows the hidden representations to grow as the data necessitates.

Neural word embeddings, such as word2vec (Mikolov et al., 2013), have become increasingly popular in both academic and industrial NLP. These methods attempt to capture the semantic meanings of words by processing huge unlabeled corpora with methods inspired by neural networks and the recent onset of Deep Learning. The result is a vectorial representation of every word in a low-dimensional continuous space. These word vectors exhibit interesting arithmetic properties (e.g. king – man + woman = queen) (Mikolov et al., 2013), and seemingly outperform traditional vector-space models of meaning inspired by Harris’s Distributional Hypothesis (Baroni et al., 2014). Our work attempts to demystify word embeddings, and understand what makes them so much better than traditional methods at capturing semantic properties.

Our main result shows that state-of-the-art word embeddings are actually “more of the same”. In particular, we show that skip-grams with negative sampling, the latest algorithm in word2vec, is implicitly factorizing a word-context PMI matrix, which has been thoroughly used and studied in the NLP community for the past 20 years. We also identify that the root of word2vec’s perceived superiority can be attributed to a collection of hyperparameter settings. While these hyperparameters were thought to be unique to neural-network inspired embedding methods, we show that they can, in fact, be ported to traditional distributional methods, significantly improving their performance. Among our qualitative results is a method for interpreting these seemingly-opaque word-vectors, and the answer to why king – man + woman = queen.

Bio: Omer Levy is a post-doc in the Department of Computer Science & Engineering at the University of Washington, working with Prof. Luke Zettlemoyer. Previously, he completed his BSc and MSc at Technion – Israel Institute of Technology with the guidance of Prof. Shaul Markovitch, and got his PhD at Bar-Ilan University with the supervision of Prof. Ido Dagan and Dr. Yoav Goldberg. Omer is interested in realizing high-level semantic applications such as question answering and summarization to help people cope with information overload. At the heart of these applications are challenges in textual entailment, semantic similarity, and reading comprehension, which form the core of my current research. He is also interested in the current advances in deep learning and how they can facilitate semantic applications.

Duolingo is a language education platform that teaches 20 languages to more than 150 million students worldwide. Our free flagship learning app is the \#1 way to learn a language online, and is the most-downloaded education app for both Android and iOS devices. In this talk, I will describe the Duolingo system and several of our empirical research projects to date, which combine machine learning with computational linguistics and psychometrics to improve learning, engagement, and even language proficiency assessment through our products.

Convolutional Neural Net (CNN) architectures have terrific recognition performance but rely on spatial pooling which makes it difficult to adapt them to tasks that require dense, pixel-accurate labeling. We make two contributions to solving this problem: (1) We demonstrate that while the apparent spatial resolution of convolutional feature maps is low, the high-dimensional feature representation contains significant sub-pixel localization information. (2) We describe a multi-resolution reconstruction architecture based on a Laplacian pyramid that uses skip connections from higher resolution feature maps and multiplicative gating to successively refine segment boundaries reconstructed from lower-resolution maps. This approach yields state-of-the-art semantic segmentation results on the PASCAL VOC and Cityscapes segmentation benchmarks without resorting to more complex random-field inference or instance detection driven architectures.

Despite the rapid development of computer graphics during the recent years, complex materials such as fabrics, fur, and human hair remain largely lacking in the virtual worlds. This is due to both the lack of high-fidelity data and the inability to efficiently describe these complicated objects via mathematical/statistical models.

In this talk, I will present my research that introduces new means to acquire, model, and render complex materials that are essential to our daily lives with a focus on fabrics. Leveraging detailed geometric information and sophisticated optical model, our work has led to computer generated imagery with a new level of accuracy and fidelity. In particular, we measure real-world samples using volume imaging (e.g., computed micro-tomography) to obtain detailed datasets on their micro-geometries. We then fit sophisticated statistical models to the measured data, yielding highly compact yet realistic representations. Lastly, we show how to recover a sample’s optical properties (e.g., colors) using optimization.

Recent technological developments are creating new spatio-temporal data streams that contain a wealth of information relevant to sustainable development goals. Modern AI techniques have the potential to yield accurate, inexpensive, and highly scalable models to inform research and policy. As a first example, I will present a machine learning method we developed to predict and map poverty in developing countries. Our method can reliably predict economic well-being using only high-resolution satellite imagery. Because images are passively collected in every corner of the world, our method can provide timely and accurate measurements in a very scalable end economic way, and could revolutionize efforts towards global poverty eradication. As a second example, I will present some ongoing work on monitoring agricultural and food security outcomes from space.

This talks explores recent uses of machine learning to large proprietary consumer transaction datasets. These are datasets which record barcode level transaction information on individual items purchased grouped by shopping trip and customer. Recent innovations in data collection allow us to go beyond the supermarket scanner to collect such data and include recent efforts to digitize the universe of customers’ receipts across all channels from supermarkets to online purchases. Additionally, passive wifi tracking allows us to record search behavior in stores and model how it translates into sales. It also gives us the opportunity to create real time interventions to nudge consumer shopping behavior. We will explore some of the challenges of modeling consumer behavior using these data and discuss methods such as tensor decompositions for count data, discrete choice modeling with Dirichlet Process Mixtures, and the use of deep autoencoders for producing interpretable statistical hypotheses.

This talk investigates the restricted Boltzmann machine (RBM), which is the building block for many deep probabilistic models. We propose an infinite RBM model, whose maximum likelihood estimation corresponds to a constrained convex optimization. We consider the Frank-Wolfe algorithm to solve the program, which provides a sparse solution that can be interpreted as inserting a hidden unit at each iteration. As a side benefit, this can be used to easily and efficiently identify an appropriate number of hidden units during the optimization. We also investigate different learning algorithms for conditional RBMs. There is a pervasive opinion that loopy belief propagation does not work well on RBM-based models, especially for learning. We demonstrate that, in the conditional setting, learning RBM-based models with belief propagation and its variants can provide much better results than the state-of-the-art contrastive divergence algorithms.

Traditionally, the field of computational Bayesian statistics has been divided into two main subfields: variational inference and Markov chain Monte Carlo (MCMC). In recent years, however, several methods have been proposed based on combining variational Bayesian inference and MCMC simulation in order to improve their overall accuracy and computational efficiency. This marriage of fast evaluation and flexible approximation provides a promising means of designing scalable Bayesian inference methods. In this work, we explore the possibility of incorporating variational approximation into a state-of-the-art MCMC method, Hamiltonian Monte Carlo (HMC), to reduce the required expensive computation involved in the sampling procedure, which is the bottleneck for many applications of HMC in big data problems. To this end, we exploit the regularity in parameter space to construct a free-form approximation of the target distribution by a fast and flexible surrogate function using an optimized additive model of proper random basis. The surrogate provides sufficiently accurate approximation while allowing for fast computation, resulting in an efficient approximate inference algorithm. We demonstrate the advantages of our method on both synthetic and real data problems.

Many machine learning problems, especially scientific problems in areas such as ecology, climate science, and brain sciences, operate in the so-called `low samples, high dimensions’ regime. Such problems typically have numerous possible predictors or features, but the number of training examples is small, often much smaller than the number of features. In this talk, we will discuss recent advances in general formulations and estimators for such problems. These formulations generalize prior work such as the Lasso and the Dantzig selector. We will discuss the geometry underlying such formulations, and how the geometry helps in establishing finite sample properties of the estimators. We will also discuss applications of such results in structure learning in probabilistic graphical models, along with real world applications in ecology and climate science.

This is joint work with Soumyadeep Chatterjee, Sheng Chen, Farideh Fazayeli, Andre Goncalves, Jens Kattge, Igor Melnyk, Peter Reich, Franziska Schrodt, Hanhuai Shan, and Vidyashankar Sivakumar.

Stein’s method provides a remarkable theoretical tool in probability theory but has not been widely known or used in practical machine learning. In this talk, we try to bright this gap and show that some of the key ideas of Stein’s method can be naturally combined with practical machine learning and probabilistic inference techniques such as kernel method, variational inference and variance reduction, which together form a new general framework for deriving new algorithms for handling the kind of highly complex, structured probabilistic models widely used in modern (deep) machine learning. The new algorithms derived in this way often have a simple, untraditional form and have significant advantages over the traditional methods. I will show several applications, including goodness-of-fit tests for evaluating models without knowing the normalization constants, scalable Bayesian inference that combines the advantages of variational inference, Monte Carlo and gradient-based optimization, and approximate maximum likelihood training of deep generative models that can generate realistic-looking images.

We develop upper and lower bounds for the probability of Boolean functions by treating multiple occurrences of variables as independent and assigning them new individual probabilities. We call this approach “dissociation” and give an exact characterization of optimal oblivious bounds, i.e. when the new probabilities are chosen independent of the probabilities of all other variables.

Our motivation comes from the weighted model counting problem (or, equivalently, the problem of computing the probability of a Boolean function), which is \#P-hard in general. By performing several dissociations, one can transform a Boolean formula whose probability is difficult to compute, into one whose probability is easy to compute, and which is guaranteed to provide an upper or lower bound on the probability of the original formula by choosing appropriate probabilities for the dissociated variables. Our new bounds shed light on the connection between previous relaxation-based and model-based approximations and unify them as concrete choices in a larger design space. We also show how our theory allows a standard relational database management system to both upper and lower bound hard probabilistic queries in guaranteed polynomial time. (Based on joint work with Dan Suciu from TODS 2014, VLDB 2015, and VLDBJ 2016: http://arxiv.org/pdf/1409.6052,http://arxiv.org/pdf/1412.1069, http://arxiv.org/pdf/1310.6257)

Extracting structured planted subgraphs from large graphs is a fundamental question that arises in a range of application domains. We describe a computationally tractable approach based on convex optimization to recover certain families of structured graphs that are embedded in larger graphs containing spurious edges. Our method relies on tractable semidefinite descriptions of majorization inequalities on the spectrum of a matrix, and we give conditions on the eigenstructure of a planted graph in relation to the noise level under which our algorithm succeeds. (Joint work with Utkan Candogan.)

Clinical medical data, especially in the intensive care unit (ICU), consist of multivariate time series of observations. For each patient visit (or episode), sensor data and lab test results are recorded in the patient’s Electronic Health Record (EHR). While potentially containing a wealth of insights, the data is difficult to mine effectively, owing to varying length, irregular sampling and missing data. Recurrent Neural Networks (RNNs), particularly those using Long Short-Term Memory (LSTM) hidden units, are powerful and increasingly popular models for learning from sequence data. They effectively model varying length sequences and capture long range dependencies. We present the first study to empirically evaluate the ability of LSTMs to recognize patterns in multivariate time series of clinical measurements. Specifically, we consider multilabel classification of diagnoses, training a model to classify 128 diagnoses given 13 frequently but irregularly sampled clinical measurements. First, we establish the effectiveness of a simple LSTM network for modeling clinical data. Then we demonstrate a straightforward and effective training strategy in which we replicate targets at each sequence step. Trained only on raw time series, our models outperform several strong baselines, including a multilayer perceptron trained on hand-engineered features.

Bayesian inference has found widespread application and use in science and engineering to reconcile Earth system models with data, including prediction in space (interpolation), prediction in time (forecasting), assimilation of observations and deterministic/stochastic model output, and inference of the model parameters. In this talk I will review the basic elements of the DiffeRential Evolution Adaptive Metropolis (DREAM) algorithm developed by Vrugt et al. (2008,2009) and used for Bayesian inference in fields ranging from physics, chemistry and engineering, to ecology, hydrology, and geophysics. I will also discuss recent developments of DREAM, including the development of a diagnostic model evaluation framework using likelihood free inference, and the use of dimensionality reduction techniques for calibration of CPU-intensive system models. Practical examples are used from many different fields of study.

Humans can describe observations and act upon requests. This requires that language be grounded in perception and motor control. I will present several components of my long-term research program to understand the vision-language-motor interface in the human brain and emulate such on computers.

In the first half of the talk, I will present fMRI investigation of the vision-language interface in the human brain. Subjects were presented with stimuli in different modalities—spoken sentences, textual presentation of sentences, and video clips depicting activity that can be described by sentences—while undergoing fMRI. The scan data is analyzed to allow readout of individual constituent concepts and words—people/names, objects/nouns, actions/verbs, and spatial-relations/prepositions—as well as phrases and entire sentences. This can be done across subjects and across modality; we use classifiers trained on scan data for one subject to read out from another subject and use classifiers trained on scan data for one modality, say text, to read out from scans of another modality, say video or speech. Analysis of this indicates that the brain regions involved in processing the different kinds of constituents are largely disjoint but also largely shared across subjects and modality. Further, we can determine the predication relations; when the stimuli depict multiple people, objects, and actions, we can read out which people are performing which actions with which objects. This points to a compositional mental semantic representation common across subjects and modalities.

In the second half of the talk, I will use this work to motivate the development of three computational systems. First, I will present a system that can use sentential description of human interaction with previously unseen objects in video to automatically find and track those objects. This is done without any annotation of the objects and without any pretrained object detectors. Second, I will present a system that learns the meanings of nouns and prepositions from video and tracks of a mobile robot navigating through its environment paired with sentential descriptions of such activity. Such a learned language model then supports both generation of sentential description of new paths driven in new environments as well as automatic driving of paths to satisfy navigational instructions specified with new sentences in new environments. Third, I will present a system that can play a physically grounded game of checkers using vision to determine game state and robotic arms to change the game state by reading the game rules from natural-language instructions.

Joint work with Andrei Barbu, Daniel Paul Barrett, Charles Roger Bradley, Seth Benjamin Scott Alan Bronikowski, Zachary Burchill, Wei Chen, N. Siddharth, Caiming Xiong, Haonan Yu, Jason J. Corso, Christiane D. Fellbaum, Catherine Hanson, Stephen Jose Hanson, Sebastien Helie, Evguenia Malaia, Barak A. Pearlmutter, Thomas Michael Talavage, and Ronnie B. Wilbur.

Bio: Jeffrey M. Siskind received the B.A. degree in computer science from the Technion, Israel Institute of Technology, Haifa, in 1979, the S.M. degree in computer science from the Massachusetts Institute of Technology (M.I.T.), Cambridge, in 1989, and the Ph.D. degree in computer science from M.I.T. in 1992. He did a postdoctoral fellowship at the University of Pennsylvania Institute for Research in Cognitive Science from 1992 to 1993. He was an assistant professor at the University of Toronto Department of Computer Science from 1993 to 1995, a senior lecturer at the Technion Department of Electrical Engineering in 1996, a visiting assistant professor at the University of Vermont Department of Computer Science and Electrical Engineering from 1996 to 1997, and a research scientist at NEC Research Institute, Inc. from 1997 to 2001. He joined the Purdue University School of Electrical and Computer Engineering in 2002 where he is currently an associate professor. His research interests include computer vision, robotics, artificial intelligence, neuroscience, cognitive science, computational linguistics, child language acquisition, automatic differentiation, and programming languages and compilers.

Circadian rhythms date back to the origins of life, are found in virtually every species and every cell, and play fundamental roles in functions ranging from metabolism to cognition. Modern high-throughput technologies allow the measurement of concentrations of transcripts, metabolites, and other species along the circadian cycle creating novel computational challenges and opportunities, including the problems of inferring whether a given species oscillate in circadian fashion or not, and inferring the time at which a set of measurements was taken. Due to the expensive process of taking these measurements, inferring whether a given species oscillate in circadian fashion has proven to be a challenge. The sparse data with only a few replicates makes many existing methods unreliable. In addition, many differential gene expression experiments–such as those contained in the GEO repository, have been carried at single time points without taking into account circadian oscillations which can act as confounding factors. To solve these problems we introduce two deep learning methods: BIO_CYCLE and BIO_CLOCK. BIO_CYCLE takes advantage of synthetic data to determine whether or not a signal oscillates in a circadian fashion, and infer periods, amplitudes, and phases. BIO_CLOCK, using a specialized cost function and real-world data, imputes the time at which a sample was taken, from the corresponding gene expression measurements. These tools are a necessary step forward to better understand circadian rhythms at the molecular level and their applications to precision medicine.

In this talk, I’ll present my group’s work on addressing two key challenges in developing parallel algorithms and software for the class of N-body problems on current and future platforms. The first challenge is reducing the apparent gap in performance between code generated from high-level forms and that of hand-tuned code, which we address using extensive characterization of the optimization space for these computations and automating the process through domain specific code generators. These application-specific compilers provide the domain scientists the ability to productively harness the power of these large machines and to enable large-scale scientific simulations and big data applications.

The second challenge is analyzing and designing algorithms. We are entering the era of exascale. The number of cores are growing at a much faster rate than bandwidth per node. What implications does this trend have in designing algorithms for future systems? If we were to model computation and communication costs, what inferences can we derive from such a model for the time to execute an algorithm? Our model suggests a new kind of high level analytical co-design of the algorithm and architecture and similar analysis can be applied in designing algorithms in general.

Several problems arising in the design, analysis and efficient operation of power systems are naturally posed as graph-structured optimization problems. Due to the nonlinear nature of the physical equations describing the power grid, these problems are often nonconvex and NP-hard. However, practical instances of several graph-structured optimization problems have been solved successfully in the graphical models literature by exploiting graph structure and using message-passing or belief propagation techniques. In this work, we show that a similar approach can be successfully applied to power systems, leading to theoretically and practically efficient algorithms. I will discuss two applications in detail: a) Solving mixed-integer optimal power flow problems on distribution networks, and b) Detecting and mitigating market manipulation by aggregators of renewable generation in a distribution-level market. I will also discuss possible extensions of these approaches to other power system/infrastructure network problems.
Based on joint work with Misha Chertkov, Sidhant Misra, Marc Vuffray, Pascal Van Hentenryck, Niangjun Chen, Navid Azizan Ruhi and Adam Wierman.

Social network analysis has a long and successful history in the social sciences, often with a focus on relatively small survey-based data sets. In the past decade, driven by the ease of automatically collecting large-scale network data sets, there has been significant interest in developing new statistical and machine learning techniques for network analysis. In this talk we will focus on two general modeling themes in this context: the use of latent variables for low-dimensional vector-based network representations models and event-based models for temporal network data. We will review the representational capabilities of these models from a generative perspective, discuss some of the challenges of parameter estimation that arise, and emphasize the role of predictive evaluation. The talk will conclude with a brief discussion of future directions in this general area.

Based on joint work with Zach Butler, Chris DuBois, Jimmy Foulds, and Carter Butts

Topic models have become increasingly prominent text-analytic machine learning tools for research in the social sciences and the humanities. In particular, custom topic models can be developed to answer specific research questions. The design of these models requires a nontrivial amount of effort and expertise, motivating general-purpose topic modeling frameworks. In this talk I will introduce latent topic networks, a flexible class of richly structured topic models designed to facilitate applied research. Custom models can straightforwardly be developed in this framework with an intuitive first-order logical probabilistic programming language. Latent topic networks admit scalable training via a parallelizable EM algorithm which leverages ADMM in the M-step. I demonstrate the broad applicability of the models with case studies on modeling influence in citation networks, and U.S. Presidential State of the Union addresses. This talk is based on joint work with Lise Getoor and Shachi Kumar from the University of California, Santa Cruz, published at ICML 2015.

Latent or hidden variable models have applications in almost every domain, e.g., social network analysis, natural language processing, computer vision and computational biology. Training latent variable models is challenging due to non-convexity of the likelihood objective function. An alternative method is based on the spectral decomposition of low order moment matrices and tensors. This versatile framework is guaranteed to estimate the correct model consistently. I will discuss my results on convergence to globally optimal solution for stochastic gradient descent, despite non-convexity of the objective. I will then discuss large-scale implementations (which are highly parallel and scalable) of spectral methods, carried out on CPU/GPU and Spark platforms. We obtain a gain in both accuracies and in running times by several orders of magnitude compared to the state-of-art variational methods. I will discuss the following applications in detail: (1) learning hidden user commonalities (communities) in social networks, and (2) learning sentence embeddings for paraphrase detection using convolutional models. More generally, I have applied the methods to a variety of problems such as text and social network analysis, healthcare analytics, and cataloging neuronal cell types in neuroscience.

Optimization lies at the core of machine learning. However, most machine learning problems entail non-convex optimization. In this talk, I will show how spectral and tensor methods can yield guaranteed convergence to globally optimal solutions under transparent conditions for a range of machine learning problems.

In the first part, I will explain how tensor methods are useful for learning latent variable models in an unsupervised manner. The focus of my work is on overcomplete regime where the hidden dimension is larger than the observed dimensionality. I describe how tensor methods enable us to learn these models in the overcomplete regime with theoretical guarantees in recovering the parameters of the model. I also provide efficient sample complexity results for training these models. Next, I will describe a new method for training neural networks for which we provide theoretical guarantees on the performance of the algorithm. We have developed a computationally efficient algorithm for training a two-layer neural network using method-of-moment and tensor decomposition techniques.

I will discuss subsampling and sketching with their applications and analysis in machine learning. They can be viewed not only as tools to improve computational and storage efficiency of existing learning algorithms, but also as settings that characterize data measurement/availability/privacy constraints in modern machine learning applications. In this talk I will introduce my recent work, which analyze subsampling and sketching settings in three popular machine learning algorithms: tensor factorization, subspace clustering and linear regression.

Understanding the semantics of preferences and behavior is incredibly complicated, especially in settings where the visual appearance of items influences our decisions. Three challenges that I’ll discuss in this talk include (1) how can we uncover the semantics of visual preferences, especially in sparse or long-tailed data, where new items are constantly introduced? (2) How can we use visual data to understand the relationships between items, and in particular what makes two items “visually compatible”? And (3) how can we understand the temporal dynamics of visual preferences, in order to uncover how “fashions” have evolved over time?

We investigate the potential of look-ahead in the context of AND/OR search in graphical models using the mini-bucket heuristic for combinatorial optimization tasks (e.g. MAP/MPE or weighted CSPs.) We present and analyze the complexity of computing the residual (a.k.a. Bellman update) of the mini-bucket heuristic, which we call “bucket errors” and show how this can be used to identify which parts of the search space are more likely to benefit from look-ahead, therefore facilitating a method to bound its overhead. We also rephrase the look-ahead computation as a graphical model to make use of structure exploiting inference schemes. In our empirical results, we demonstrate that our methods can be used to cost-effectively increase the power of branch-and-bound search.

In the second part of the talk, we show how bucket errors can be used to improve the performance of AND/OR best-first search algorithms for providing lower bounds on the min-sum problem. In our preliminary experiments, we show that when expanding nodes for the AO* algorithm, using bucket errors as a subproblem ordering heuristic can allow us to expand fewer nodes to arrive at the optimal solution compared to the existing ordering approach.

Learning with big data is a challenging task that requires smart and efficient methods to extract useful information from data. Optimization methods, both convex and nonconvex are promising approaches to do this. In this talk I will review two classes of my work on prominent problems in convex and nonconvex optimization.

Beating the Perils of Non-Convexity: Guaranteed Training of Neural Networks using Tensor Method: Neural networks provide a versatile tool for approximating functions of various inputs. Despite exciting achievements in application, a theoretical understanding of them is mostly lacking. Training a neural network is a highly nonconvex problem and backpropagation can get stuck in local optima. For the first time, we have a computationally efficient method for training neural networks that also has guaranteed generalization. This is part of our recently proposed general framework based on method-of-moments and tensor decomposition to efficiently learn different models such as neural networks and mixture of classifiers.

Breaking Curse of Dimensionality: Stochastic Optimization in high dimensions: We have designed an efficient stochastic optimization method based on ADMM that is fast and cheap to implement, can be performed in parallel and can be used for any regularized optimization framework with some mild assumptions. We have proved that our algorithm obtains minimax optimal convergence rates for sparse optimization and robust PCA framework. Experiment results show that in the aforementioned scenarios, our method outperforms state-of-the-art, i.e., yields smaller error with equal time.

We are proposing an algorithm to test the accuracy of the predictions by earthquake early warning systems. Most warning systems predict the location and the magnitude of an ongoing earthquake via the early-arriving seismic wave data. Our algorithm uses logarithm of ratios between observed ground motion envelopes and Virtual Seismologist’s (Cua G. and Heaton T.) predicted envelopes to assess the validity of system predictions. We quantify the uncertainty attached to our parameters using Bayesian probability approach.

Importance sampling (IS) and its variant, annealed IS (AIS) have been widely used for estimating the partition function in graphical models, such as Markov random fields and deep generative models. However, IS tends to underestimate the partition function and is subject to high variance when the proposal distribution is more peaked than the target distribution. On the other hand, “reverse” versions of IS and AIS tend to overestimate the partition function, and degenerate when the target distribution is more peaked than the proposal distribution. We present a simple, general method that gives much more reliable and robust estimates than either IS (AIS) or reverse IS (AIS). Our method works by converting the estimation problem into a simple classification problem that discriminates between the samples drawn from the target and the proposal. We give both theoretical and empirical justification, and show that an annealed version of our method significantly outperforms both AIS and reverse AIS (Burda et al., 2015), which has been the state-of-the-art for likelihood evaluation in deep generative models. Joint work with Qiang Liu, Jian Peng, and John Fisher.

In this talk, we propose the multi-version coding problem for distributed storage. We consider a setting where there are n servers that aim to store v versions of a message, and there is a total ordering on the versions from the earliest to the latest. We assume that each message version has a given number of bits. Each server can receive any subset of the v versions and stores a function of the message versions it receives. The multi-version code we consider ensures that, a decoder that connects to any c out of the n servers can recover the message corresponding to the latest common version stored among those servers, or a message corresponding to a version that is later than the latest common version. We describe a simple and explicit achievable scheme, as well as an information-theoretic converse. Moreover, we apply the multi-version code to one of the problems in distributed algorithms – the emulation of atomic shared memory in a message-passing network – and improve upon previous algorithms up to a half in terms of storage cost.

Network models provide a unifying framework for understanding dependencies among variables in medical, biological, and other sciences. Networks can be used to reveal underlying data structures, infer functional modules, and facilitate experiment design. In practice, however, size, uncertainty and complexity of the underlying associations render these applications challenging.

In this talk, we illustrate the use of spectral, combinatorial, and statistical inference techniques in several significant network science problems. First, we consider the problem of network alignment where the goal is to find a bijective mapping between nodes of two networks to maximize their overlapping edges while minimizing mismatches. To solve this combinatorial problem, we present a new scalable spectral algorithm, and establish its efficiency theoretically and experimentally over several synthetic and real networks. Next, we introduce network maximal correlation (NMC) as an essential measure to capture nonlinear associations in networks. We characterize NMC using geometric properties of Hilbert spaces and illustrate its application in learning network topology when variables have unknown nonlinear dependencies. Finally, we discuss the problem of learning low dimensional structures (such as clusters) in large networks, where we introduce logistic Random Dot Product Graphs, a new class of networks which includes most stochastic block models as well as other low dimensional structures. Using this model, we propose a spectral network clustering algorithm that possesses robust performance under different clustering setups. In all of these problems, we examine underlying fundamental limits and present efficient algorithms for solving them. We also highlight applications of the proposed algorithms to data-driven problems such as functional and regulatory genomics of human diseases, and cancer.

Bio: Soheil Feizi is a PhD candidate at Massachusetts Institute of Technology (MIT), co-supervised by Prof. Muriel Médard and Prof. Manolis Kellis. His research interests include analysis of complex networks and the development of inference and learning methods based on Optimization, Information Theory, Machine Learning, Statistics, and Probability, with applications in Computational Biology, and beyond. He completed his B.Sc. at Sharif University of Technology, awarded as the best student of his class. He received the Jacobs Presidential Fellowship and EECS Great Educators Fellowship, both from MIT. He has been a finalist in the Qualcomm Innovation contest. He received an Ernst Guillemin Award for his Master of Science Thesis in the department of Electrical Engineering and Computer Science at MIT.

Neuronal networks have enjoyed a resurgence both in the worlds of neuroscience, where they yield mathematical frameworks for thinking about complex neural datasets, and in machine learning, where they achieve state of the art results on a variety of tasks, including machine vision, speech recognition, and language translation. Despite their empirical success, a mathematical theory of how deep neural circuits, with many layers of cascaded nonlinearities, learn and compute remains elusive. We will discuss three recent vignettes in which ideas from statistical physics can shed light on this issue. In particular, we show how dynamical criticality can help in neural learning, how the non-intuitive geometry of high dimensional error landscapes can be exploited to speed up learning, and how modern ideas from non-equilibrium statistical physics, like the Jarzynski equality, can be extended to yield powerful algorithms for modeling complex probability distributions. Time permitting, we will also discuss the relationship between neural network learning dynamics and the developmental time course of semantic concepts in infants.

Weighted Max-SAT is an extension of SAT in which each clause has an associated cost. The goal is to minimize the cost of falsified clauses. Max-SAT has been successfully applied to a number of domains including Bioinformatics, Telecommunications and Scheduling.

In this talk I will introduce the Max-SAT framework and discuss the main solving approaches. In particular, I will present Max-resolution and will show how it can be effectively used in the context of Depth-first Branch-and-Bound.

Occlusion poses a significant difficulty for detecting and localizing object keypoints and subsequent fine-grained identification. In this talk, I will describe a hierarchical deformable part model for face detection and keypoint localization that explicitly models part occlusion. The proposed model structure makes it possible to augment positive training data with large numbers of synthetically occluded instances. This allows us to easily incorporate the statistics of occlusion patterns in a discriminatively trained model. However, this model does not exploit bottom-up cues such as detection of occluding contours and image segments. I will talk about how to modify the proposed model to utilize bottom-up class-specific segmentation in order to jointly detect and segment out the foreground pixels belonging to the face.

In many classification problems labels are relatively scarce. One context in which this occurs is where we have labels for groups of instances but not for the instances themselves, as in multi-instance learning. Past work on this problem has typically focused on learning classifiers to make predictions at the group level. In this paper we focus on the problem of learning classifiers to make predictions at the instance level. To achieve this we propose a new objective function that encourages smoothness of inferred instance-level labels based on instance-level similarity, while at the same time respecting group-level label constraints. We apply this approach to the problem of predicting labels for sentences given labels for reviews, using a convolutional neural network to infer sentence similarity. The approach is evaluated using three large review data sets from IMDB, Yelp, and Amazon, and we demonstrate the proposed approach is both accurate and scalable compared to various alternatives.