Machine learning is a powerful tool for tackling challenging problems
in artificial intelligence. In practice, success of machine learning
algorithms critically depends on the feature representations for input
data, which often becomes a limiting factor. To address this problem,
deep learning methods have recently emerged as successful techniques
to learn feature hierarchies from unlabeled and labeled data. In this
talk, I will present my perspectives on the progress, challenges, and
some new directions. Specifically, I will talk about my recent work to
address the following interrelated challenges: (1) how can we learn
invariant yet discriminative features, and furthermore disentangle
underlying factors of variation to model high-order interactions
between the factors? (2) how can we learn representations of the
output data when the output variables have complex high-order
dependencies? (3) how can we learn shared representations from
heterogeneous input data modalities?
Honglak Lee is an Assistant Professor of Computer Science and
Engineering at the University of Michigan, Ann Arbor. He received his
Ph.D. from Computer Science Department at Stanford University in 2010,
advised by Prof. Andrew Ng. His primary research interests lie in
machine learning, which spans over deep learning, unsupervised and
semi-supervised learning, transfer learning, graphical models, and
optimization. He also works on application problems in computer
vision, audio recognition, robot perception, and text processing. His
work received best paper awards at ICML and CEAS. He has served as a
guest editor of IEEE TPAMI Special Issue on Learning Deep
Architectures, as well as area chairs of ICML and NIPS. He received
the Google Faculty Research Award in 2011, and was selected by IEEE
Intelligent Systems as one of AI's 10 to Watch in 2013.
Amazon has many applications whose core is multilabel
classification. This talk will present progress towards a multilabel
learning method that can handle 10^7 training examples, 10^6 features, and
10^5 labels on a single workstation. A sparse linear model is learned for
each label simultaneously by stochastic gradient descent with L2 and L1
regularization. Tractability is achieved through careful use of sparse data
structures, and speed is achieved by using the latest stochastic gradient
methods that do variance reduction. Both theoretically and practically,
these methods achieve order-of-magnitude faster convergence than Adagrad.
We have extended them to handle non-differentiable L1 regularization. We
show experimental results on classifying biomedical articles into 26,853
scientific categories. [Joint work with Galen Andrew, ML intern at Amazon.]
Bio Charles Elkan is the first Amazon Fellow, on leave from being a
professor of computer science at the University of California, San Diego.
In the past, he has been a visiting associate professor at Harvard and a
researcher at MIT. His published research has been mainly in machine
learning, data science, and computational biology. The MEME algorithm that
he developed with Ph.D. students has been used in over 3000 published
research projects in biology and computer science. He is fortunate to have
had inspiring undergraduate and graduate students who are in leadership
positions now such as vice president at Google.
Probabilistic modeling of ranking data is an extensively studied
problem with applications ranging from understanding user preferences
in electoral systems and social choice theory, to more modern learning
tasks in online web search, crowd-sourcing and recommendation
systems. This work concerns learning the Mallows model -- one of the
most popular probabilistic models for analyzing ranking data. In this
model, the user's preference ranking is generated as a noisy version
of an unknown central base ranking. The learning task is to recover
the base ranking and the model parameters using access to noisy
rankings generated from the model.
Although well understood in the setting of a homogeneous population (a
single base ranking), the case of a heterogeneous population (mixture
of multiple base rankings) has so far resisted algorithms with
guarantees on worst case instances. In this talk I will present the
first polynomial time algorithm which provably learns the parameters
and the unknown base rankings of a mixture of two Mallows models. A
key component of our algorithm is a novel use of tensor decomposition
techniques to learn the top-k prefix in both the rankings. Before this
work, even the question of identifiability in the case of a mixture of
two Mallows models was unresolved.
Joint work with Avrim Blum, Or Sheffet and Aravindan Vijayaraghavan.
Many applications involve multiple interlinked data sources, but existing
approach to handle them are often based on latent factor models (i.e.
distributed representations) which are difficult to learn. At the same
time, recent advances in convex analysis, mainly based on the nuclear norm
(relaxation of the matrix rank) and sparse structured approximations, have
shown great theoretical and practical performances to handle very large
matrix factorization problems with non-Gaussian noise and missing data.
In this talk, we will show how multiple matrices or tensors can be jointly
factorized using a convex formulation of the problem, with a particular
- Multi-view learning: A popular approach is to assume that, both, the
correlations between the views and the view-specific correlations have
low-rank structure, leading to a model closely related to canonical
correlation analysis called inter-battery factor analysis. We propose a
convex relaxation of this model, based on a structured nuclear norm
- Collective matrix factorization: When multiple matrices are related, they
share common latent factors, leading to a simple yet powerful way of
handling complex data structures, such as relational databases. Again, a
convex formulation of this approach is proposed. We also show that the
Bayesian version of this model can be used to tune the multiple
regularization parameters involved in such models, avoiding costly
Another contribution to KB modeling relates to binary tensor and matrix
factorization with many zeros. We show a new learning approaches for binary
data that scales linearly with the number of positive examples. It is based
on a iterative split of the tensor (or matrix) on which the binary loss is
approximated by a Gaussian loss which itself can be efficiently minimized.
Experiments on popular tasks such as data imputation, multi-label
prediction, link prediction in graphs and item recommendation illustrate
the benefit of the proposed approaches.
BioGuillaume Bouchard is senior research scientist at Xerox Research
Centre Europe in Grenoble, France. After an engineering degree and master
in mathematics in Université de Rouen, he obtained a PhD in statistics from
Institut National de Recherche en Information et Automatique (INRIA) in
2004. Since then, he worked for Xerox on multiple machine learning research
project in big data analysis, including user modelling, recommender systems
and natural language processing. He was involved in French and European
research projects called LAVA, FUPOL, Fusepool and Dynamicité. His current
research focuses on the development of distributed statistical relational
models for knowledge bases, applied to the development of virtual agents.
Forward stagewise regression follows a very simple strategy for constructing a sequence of sparse regression estimates: starting with all coefficients equal to zero, it iteratively updates the coefficient (by a small amount ε) corresponding to the variable that has maximal absolute inner product with the current residual. This procedure has an interesting connection to the lasso: under some conditions, it can be shown that the sequence of forward stagewise estimates exactly coincides with the lasso path, as the step size ε goes to zero. Further, essentially the same equivalence holds outside of the regression setting, for minimizing a differentiable convex loss function subject to an l1 norm constraint (and the stagewise algorithm now updating the coefficient corresponding to the maximal absolute component of the gradient).
Even when they do not match their l1-constrained analogues, stagewise estimates provide a useful approximation, and are computationally appealing. Their success in sparse modeling motivates the question: can a simple, effective strategy like forward stagewise be applied more broadly in other regularization settings, beyond the l1 norm and sparsity? This is the focus of the talk; we present a general framework for stagewise estimation, which yields fast algorithms for problems such as group-structured learning, matrix completion, image denoising, and more.
Machine learning algorithms frequently involve careful tuning of learning parameters and model hyperparameters. Unfortunately, this tuning is often a "black art" requiring expert experience, rules of thumb, or sometimes brute-force search. There is therefore great appeal for automatic approaches that can optimize the performance of any given learning algorithm to the problem at hand. I will describe my recent work on solving this problem with Bayesian nonparametrics, using Gaussian processes. This approach of "Bayesian optimization" models the generalization performance as an unknown objective function with a GP prior. I will discuss new algorithms that account for variable cost in function evaluation and take advantage of parallelism in evaluation. These new algorithms improve on previous automatic procedures and can reach or surpass human expert-level optimization for many algorithms including latent Dirichlet allocation for text analysis, structured SVMs for protein motif finding, and convolutional neural networks for visual object recognition.
Within high-dimensional genomic data, including genetic markers, complex phenotypes, and gene expression measurements, hides a substantial amount of low dimensional structure. In order to extract this low dimensional structure from high dimensional data, latent factor models with appropriate priors and modeling assumptions can be applied. A key advantage of sparse latent factor models in genomic applications is interpretability: the low dimensional structure can often be understood in a biological context. Model specification here is important, however, as genomics data generally includes small numbers of samples n with respect to enormous numbers of features p. Sparse linear latent factor models have been used to identify population substructure. Applying these models to high dimensional noisy phenotypes produces interpretable phenotypes that can subsequently tested for association with genetic variants. This same idea can be considered for gene expression traits, where the low dimensional structure captures groups of interacting genes, and the associated genetic variants represent master regulators. In this talk, I will discuss recent work on sparse latent factor models, and results on genomic data.
Having long been recognized in combinatorics, the concept of submodularity is becoming increasingly important in machine learning, since it can capture intuitive and yet non-trivial interactions between variables. This expressiveness has found many applications in machine learning, particularly in understanding structured data like text, vision and speech. Submodular functions naturally capture aspects of information, coverage and diversity in maximization problems, while also modelling notions of cooperation, sparsity, complexity and economies of scale in minimization problems. In this talk, we shall consider a large class of submodular optimization problems and motivate them with several real world applications, particularly in machine learning. We shall then provide a unifying algorithmic framework for solving several of these optimization problems, and highlight the scalability and practicality of this approach. We willl also highlight novel theoretical characterizations, which provide better connections between theory and practice. We shall ground this entire talk with a large number of applications in machine learning, including image segmentation, image correspondence, image collection summarization, feature selection, and data subset selection. This talk should be self contained and shall not require prior knowledge of submodular functions and optimization.
(this is based on joint work with Jeff Bilmes, Stefanie Jegelka, Sebastian Tschiatschek and Kai Wei)
A major challenge for machine learning in the next decade is the development of methods that continuously predict and learn from a stream of data. The scale of data and the non-stationary nature of the environment make the prevalent ``learn from a batch of i.i.d. data'' paradigm of Statistical Learning inadequate. Despite the extensive literature on sequential prediction (online learning) methods, the theoretical understanding of the subject has been lacking, and the results have been obtained on a case-by-case basis. This stands in contrast to Statistical Learning, where the inherent complexities have been well-understood in the last forty years. In this talk, we focus on no-regret online learning and develop the relevant notions of complexity in a surprising parallel to Statistical Learning Theory. This non-constructive study of inherent complexity is then augmented with a recipe for developing efficient online learning algorithms via a notion of a relaxation. To demonstrate the utility of our approach, we develop a new family of randomized methods, as well as new algorithms for collaborative filtering and node prediction.