COLT 2016 Papers

We argue that the second order bounds given in Cesa-Bianchi2006, which accumulate the square of the loss of each action separately, are loose. We propose a different form of a second order bound and conjecture the it is satisfied by NormalHedge ChaudhuriFrHs2009.

The study of property elicitation is gaining ground in statistics and machine learning as a way to view and reason about the expressive power of emiprical risk minimization (ERM). Yet beyond a widening frontier of special cases, the two most fundamental questions in this area remain open: which statistics are elicitable (computable via ERM), and which loss functions elicit them? Moreover, recent work suggests a complementary line of questioning: given a statistic, how many ERM parameters are needed to compute it? We give concrete instantiations of these important questions, which have numerous applications to machine learning and related fields.

Existing vanilla algorithms for online linear optimization have O((ηR(u) + 1/η) \sqrtT) regret with respect to any competitor u, where R(u) is a 1-strongly convex regularizer and η> 0 is a tuning parameter of the algorithm. For certain decision sets and regularizers, the so-called \emphparameter-free algorithms have \widetilde O(\sqrtR(u) T) regret with respect to any competitor u. Vanilla algorithm can achieve the same bound only for a fixed competitor u known ahead of time by setting η= 1/\sqrtR(u). A drawback of both vanilla and parameter-free algorithms is that they assume that the norm of the loss vectors is bounded by a constant known to the algorithm. There exist \emphscale-free algorithms that have O((ηR(u) + 1/η) \sqrtT \max_1 \le t \le T \norm\ell_t) regret with respect to any competitor u and for any sequence of loss vector \ell_1, …, \ell_T. Parameter-free analogue of scale-free algorithms have never been designed. Is is possible to design algorithms that are simultaneously \emphparameter-free and \emphscale-free?

Radial kernels are well-suited for machine learning over general geodesic metric spaces, where pairwise distances are often the only computable quantity available. We have recently shown that geodesic exponential kernels are only positive definite for all bandwidths when the input space has strong linear properties. This negative result hints that radial kernel are perhaps not suitable over geodesic metric spaces after all. Here, however, we present evidence that large intervals of bandwidths exist where geodesic exponential kernels have high probability of being positive definite over finite datasets, while still having significant predictive power. From this we formulate conjectures on the probability of a positive definite kernel matrix for a finite random sample, depending on the geometry of the data space and the spread of the sample.

The best arm identification problem (BEST-1-ARM) is the most basic pure exploration problem in stochastic multi-armed bandits. The problem has a long history and attracted significant attention for the last decade. However, we do not yet have a complete understanding of the optimal sample complexity of the problem: The state-of-the-art algorithms achieve a sample complexity of O(\sum_i=2^n \Delta_i^-2(\lnδ^-1 + \ln\ln\Delta_i^-1)) (\Delta_i is the difference between the largest mean and the i^th mean), while the best known lower bound is Ω(\sum_i=2^n \Delta_i^-2\lnδ^-1) for general instances and Ω(∆^-2 \ln\ln ∆^-1) for the two-arm instances. We propose to study the instance-wise optimality for the BEST-1-ARM problem. Previous work has proved that it is impossible to have an instance optimal algorithm for the 2-arm problem. However, we conjecture that modulo the additive term Ω(\Delta_2^-2 \ln\ln \Delta_2^-1) (which is an upper bound and worst case lower bound for the 2-arm problem), there is an instance optimal algorithm for BEST-1-ARM. Moreover, we introduce a new quantity, called the gap entropy for a best-arm problem instance, and conjecture that it is the instance-wise lower bound. Hence, resolving this conjecture would provide a final answer to the old and basic problem.

Planning plays an important role in the broad class of decision theory. Planning has drawn much attention in recent work in the robotics and sequential decision making areas. Recently, Reinforcement Learning (RL), as an agent-environment interaction problem, has brought further attention to planning methods. Generally in RL, one can assume a generative model, e.g. graphical models, for the environment, and then the task for the RL agent is to learn the model parameters and find the optimal strategy based on these learnt parameters. Based on environment behavior, the agent can assume various types of generative models, e.g. Multi Armed Bandit for a static environment, or Markov Decision Process (MDP) for a dynamic environment. The advantage of these popular models is their simplicity, which results in tractable methods of learning the parameters and finding the optimal policy. The drawback of these models is again their simplicity: these models usually underfit and underestimate the actual environment behavior. For example, in robotics, the agent usually has noisy observations of the environment inner state and MDP is not a suitable model. More complex models like Partially Observable Markov Decision Process (POMDP) can compensate for this drawback. Fitting this model to the environment, where the partial observation is given to the agent, generally gives dramatic performance improvement, sometimes unbounded improvement, compared to MDP. In general, finding the optimal policy for the POMDP model is computationally intractable and fully non convex, even for the class of memoryless policies. The open problem is to come up with a method to find an exact or an approximate optimal stochastic memoryless policy for POMDP models.

Motivated by online advertising auctions, we consider repeated Vickrey auctions where goods of unknown value are sold sequentially and bidders only learn (potentially noisy) information about a good’s value once it is purchased. We adopt an online learning approach with bandit feedback to model this problem and derive bidding strategies for two models: stochastic and adversarial. In the stochastic model, the observed values of the goods are random variables centered around the true value of the good. In this case, logarithmic regret is achievable when competing against well behaved adversaries. In the adversarial model, the goods need not be identical. Comparing our performance against that of the best fixed bid in hindsight, we show that sublinear regret is also achievable in this case. For both the stochastic and adversarial models, we prove matching minimax lower bounds showing our strategies to be optimal up to lower-order terms. To our knowledge, this is the first complete set of strategies for bidders participating in auctions of this type.

We consider the online sparse linear regression problem, which is the problem of sequentially making predictions observing only a limited number of features in each round, to minimize regret with respect to the best sparse linear regressor, where prediction accuracy is measured by square loss. We give an \em inefficient algorithm that obtains regret bounded by \tildeO(\sqrtT) after T prediction rounds. We complement this result by showing that no algorithm running in polynomial time per iteration can achieve regret bounded by O(T^1-δ) for any constant δ> 0 unless \textsfNP ⊆\textsfBPP. This computational hardness result resolves an open problem presented in COLT 2014 (Kale, 2014) and also posed by Zolghadr et al. (2013). This hardness result holds even if the algorithm is allowed to access more features than the best sparse linear regressor up to a logarithmic factor in the dimension.

We consider the problem of prediction with expert advice when the losses of the experts have low-dimensional structure: they are restricted to an unknown d-dimensional subspace. We devise algorithms with regret bounds that are independent of the number of experts and depend only on the rank d. For the stochastic model we show a tight bound of Θ(\sqrtdT), and extend it to a setting of an approximate d subspace. For the adversarial model we show an upper bound of O(d\sqrtT) and a lower bound of Ω(\sqrtdT).

We consider the sequential decision problem known as regret minimization, or more precisely its generalization to the vectorial or multi-criteria setup called Blackwell approachability. We assume that Nature, the decision maker, or both, might have some quitting (or terminating) actions so that the stream of payoffs is constant whenever they are chosen. We call those environments “quitting games”. We characterize convex target sets \cC that are Blackwell approachable, in the sense that the decision maker has a policy ensuring that the expected average vector payoff converges to \cC at some given horizon known in advance. Moreover, we also compare these results to the cases where the horizon is not known and show that, unlike in standard online learning literature, the necessary or sufficient conditions for the anytime version of this problem are drastically different than those for the fixed horizon.

We consider the online version of the isotonic regression problem. Given a set of linearly ordered points (e.g., on the real line), the learner must predict labels sequentially at adversarially chosen positions and is evaluated by her total squared loss compared against the best isotonic (non-decreasing) function in hindsight. We survey several standard online learning algorithms and show that none of them achieve the optimal regret exponent; in fact, most of them (including Online Gradient Descent, Follow the Leader and Exponential Weights) incur linear regret. We then prove that the Exponential Weights algorithm played over a covering net of isotonic functions has a regret bounded by O\big(T^1/3 \log^2/3(T)\big) and present a matching Ω(T^1/3) lower bound on regret. We provide a computationally efficient version of this algorithm. We also analyze the noise-free case, in which the revealed labels are isotonic, and show that the bound can be improved to O(\log T) or even to O(1) (when the labels are revealed in isotonic order). Finally, we extend the analysis beyond squared loss and give bounds for entropic loss and absolute loss.

To address difficult optimization problems, convex relaxations based on semidefinite programming are now common place in many fields. Although solvable in polynomial time, large semidefinite programs tend to be computationally challenging. Over a decade ago, exploiting the fact that in many applications of interest the desired solutions are low rank, Burer and Monteiro proposed a heuristic to solve such semidefinite programs by restricting the search space to low-rank matrices. The accompanying theory does not explain the extent of the empirical success. We focus on Synchronization and Community Detection problems and provide theoretical guarantees shedding light on the remarkable efficiency of this heuristic.

We propose and analyze a family of \emphinformation processing systems, where a finite set of experts or servers are employed to extract information about a stream of incoming jobs. Each job is associated with a hidden label drawn from some prior distribution. An inspection by an expert produces a noisy outcome that depends both on the job’s hidden label and the type of the expert, and occupies the expert for a finite time duration. A decision maker’s task is to dynamically assign inspections so that the resulting outcomes can be used to accurately recover the labels of all jobs, while keeping the system stable. Among our chief motivations are applications in crowd-sourcing, diagnostics, and experiment designs, where one wishes to efficiently discover the nature of a large number of items, using a finite pool of computational resources or human agents. We focus on the \emphcapacity of such an information processing system. Given a level of accuracy guarantee, we ask how many experts are needed in order to stabilize the system, and through what inspection architecture. Our main result provides an adaptive inspection policy that is asymptotically optimal in the following sense: the ratio between the required number of experts under our policy and the theoretical optimal converges to one, as the probability of error in label recovery tends to zero.

It has long been conjectured that hypotheses spaces suitable for data that is compositional in nature, such as text or images, may be more efficiently represented with deep hierarchical networks than with shallow ones. Despite the vast empirical evidence supporting this belief, theoretical justifications to date are limited. In particular, they do not account for the locality, sharing and pooling constructs of convolutional networks, the most successful deep learning architecture to date. In this work we derive a deep network architecture based on arithmetic circuits that inherently employs locality, sharing and pooling. An equivalence between the networks and hierarchical tensor factorizations is established. We show that a shallow network corresponds to CP (rank-1) decomposition, whereas a deep network corresponds to Hierarchical Tucker decomposition. Using tools from measure theory and matrix algebra, we prove that besides a negligible set, all functions that can be implemented by a deep network of polynomial size, require exponential size in order to be realized (or even approximated) by a shallow network. Since log-space computation transforms our networks into SimNets, the result applies directly to a deep learning architecture demonstrating promising empirical performance. The construction and theory developed in this paper shed new light on various practices and ideas employed by the deep learning community.

It is well known that Sparse PCA (Sparse Principal Component Analysis) is NP-hard to solve exactly on worst-case instances. What is the complexity of solving Sparse PCA approximately? Our contributions include: \beginenumerate \item a simple and efficient algorithm that achieves an n^-1/3-approximation; \item NP-hardness of approximation to within (1-\varepsilon), for some small constant \varepsilon > 0; \item SSE-hardness of approximation to within \em any constant factor; and \item an \exp\exp\left(Ω\left(\sqrt\log \log n\right)\right) (“quasi-quasi-polynomial”) gap for the standard semidefinite program. \endenumerate

In the noisy tensor completion problem we observe m entries (whose location is chosen uniformly at random) from an unknown n_1 \times n_2 \times n_3 tensor T. We assume that T is entry-wise close to being rank r. Our goal is to fill in its missing entries using as few observations as possible. Let n = \max(n_1, n_2, n_3). We show that if m = n^3/2 r then there is a polynomial time algorithm based on the sixth level of the sum-of-squares hierarchy for completing it. Our estimate agrees with almost all of T’s entries almost exactly and works even when our observations are corrupted by noise. This is also the first algorithm for tensor completion that works in the overcomplete case when r > n, and in fact it works all the way up to r = n^3/2-ε. Our proofs are short and simple and are based on establishing a new connection between noisy tensor completion (through the language of Rademacher complexity) and the task of refuting random constant satisfaction problems. This connection seems to have gone unnoticed even in the context of matrix completion. Furthermore, we use this connection to show matching lower bounds. Our main technical result is in characterizing the Rademacher complexity of the sequence of norms that arise in the sum-of-squares relaxations to the tensor nuclear norm. These results point to an interesting new direction: Can we explore computational vs. sample complexity tradeoffs through the sum-of-squares hierarchy?

We construct a new map from a convex function to a distribution on its domain, with the property that this distribution is a multi-scale exploration of the function. We use this map to solve a decade-old open problem in adversarial bandit convex optimization by showing that the minimax regret for this problem is \tildeO(\mathrmpoly(n) \sqrtT), where n is the dimension and T the number of rounds. This bound is obtained by studying the dual Bayesian maximin regret via the information ratio analysis of Russo and Van Roy, and then using the multi-scale exploration to construct a new algorithm for the Bayesian convex bandit problem.

Strongly Rayleigh distributions are natural generalizations of product and determinantal probability distributions and satisfy the strongest form of negative dependence properties. We show that the "natural" Monte Carlo Markov Chain (MCMC) algorithm mixes rapidly in the support of a homogeneous strongly Rayleigh distribution. As a byproduct, our proof implies Markov chains can be used to efficiently generate approximate samples of a k-determinantal point process. This answers an open question raised by Deshpande and Rademacher which was studied recently by Kang, Li-Jegelka-Sra, and Rebeschini-Karbasi.

If a concept class can be represented with a certain amount of memory, can it be efficiently learned with the same amount of memory? What concepts can be efficiently learned by algorithms that extract only a few bits of information from each example? We introduce a formal framework for studying these questions, and investigate the relationship between the fundamental resources of memory or communication and the sample complexity of the learning task. We relate our memory-bounded and communication-bounded learning models to the well-studied statistical query model. This connection can be leveraged to obtain both upper and lower bounds: we show strong lower bounds on learning parity functions with bounded communication, as well as upper bounds on solving sparse linear regression problems with limited memory.