View: session overviewtalk overview
09:00 | The Power of Depth for Feedforward Neural Networks SPEAKER: unknown ABSTRACT. We show that there are simple functions on R^d, expressible by small 3-layer feedforward neural networks, which cannot be approximated by any 2-layer network, to more than a certain constant accuracy, unless its width is exponential in the dimension. The result holds for most continuous activation functions, including rectified linear units and sigmoids, and is a formal demonstration that depth -- even if increased by 1 -- can be exponentially more valuable than width for standard feedforward neural networks. Moreover, compared to related results in the context of Boolean functions, our result requires fewer assumptions, and the proof techniques and construction are very different. |
09:10 | Benefits of depth in neural networks SPEAKER: Matus Telgarsky ABSTRACT. For any positive integer k, there exist neural networks with Θ(k^3) layers, Θ(1) nodes per layer, and Θ(1) distinct parameters which can not be approximated by networks with O(k) layers unless they are exponentially large --- they must possess Ω(2^k) nodes. This result is proved here for a class of nodes termed semi-algebraic gates which includes the common choices of ReLU, maximum, indicator, and piecewise polynomial functions, therefore establishing benefits of depth against not just standard networks with ReLU gates, but also convolutional networks with ReLU and maximization gates, and boosted decision trees (in this last case with a stronger separation: Ω(2^(k^3)) total tree nodes are required). |
09:20 | On the Expressive Power of Deep Learning: A Tensor Analysis SPEAKER: unknown ABSTRACT. 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. |
09:30 | Cortical Computation via Iterative Constructions SPEAKER: unknown ABSTRACT. We study Boolean functions of an arbitrary number of input variables that can be realized by simple iterative constructions based on constant-size primitives. This restricted type of construction needs little global coordination or control and thus is a candidate for neurally feasible computation. Valiant's construction of a majority function can be realized in this manner and, as we show, can be generalized to any uniform threshold function. We study the rate of convergence, finding that while linear convergence to the correct function can be achieved for any threshold using a fixed set of primitives, for quadratic convergence, the size of the primitives must grow as the threshold approaches 0 or 1. We also study finite realizations of this process and the learnability of the functions realized. We show that the constructions realized are accurate outside a small interval near the target threshold, where the size of the construction grows as the inverse square of the interval width. This phenomenon, that errors are higher closer to thresholds (and thresholds closer to the boundary are harder to represent), is a well-known cognitive finding. |
09:40 | A Guide to Learning Arithmetic Circuits SPEAKER: Ilya Volkovich ABSTRACT. An \emph{arithmetic circuit} is a directed acyclic graph in which the operations are $\{+,\times\}$. In this paper, we exhibit several connections between learning algorithms for arithmetic circuits and other problems. In particular, we show that: \begin{itemize} \item Efficient learning algorithms for arithmetic circuit classes imply explicit exponential lower bounds. \item General circuits and formulas can be learned efficiently with membership and equivalence queries iff they can be learned efficiently with membership queries only. \item Low-query learning algorithms for certain classes of circuits imply explicit rigid matrices. \item Learning algorithms for multilinear depth-3 and depth-4 circuits must compute square roots. \end{itemize} |
10:10 | How to calculate partition functions using convex programming hierarchies: provable bounds for variational methods SPEAKER: Andrej Risteski ABSTRACT. We consider the problem of approximating partition functions for Ising models. We make use of recent tools in combinatorial optimization: the Sherali-Adams and Lasserre convex programming hierarchies, in combination with variational methods to get algorithms for calculating partition functions in these families. These techniques give new, non-trivial approximation guarantees for the partition function beyond the regime of correlation decay. They also generalize some classical results from statistical physics about the Curie-Weiss ferromagnetic Ising model, as well as provide a partition function counterpart of classical results about max-cut on dense graphs (\cite{arora1995polynomial}). With this, we connect techniques from two areas that have apparently been disparate up to now -- optimization and counting/partition function approximations. (i.e. #-P type of problems). Furthermore, we design to the best of our knowledge the first provable, convex variational methods. Though in the literature there are a host of convex versions of variational methods (\cite{wainwright2003tree, wainwright2005new, heskes2006convexity, meshi2009convexifying}), they come with no guarantees (apart from some extremely special cases, like e.g. the graph has a single cycle \cite{weiss2000correctness}). We consider dense and expander-like graphs, and interestingly, the reason our approach works on these types of graphs is because local correlations propagate to global correlations -- completely the opposite of algorithms based on correlation decay. In the process we design novel entropy approximations based on the low-order moments of a distribution. |
10:30 | Noisy Tensor Completion via the Sum-of-Squares Hierarchy SPEAKER: unknown ABSTRACT. 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 $r = n^{3/2-\epsilon}$. 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? |
10:50 | Basis Learning as an Algorithmic Primitive SPEAKER: unknown ABSTRACT. A number of important problems in theoretical computer science and machine learning can be interpreted as recovering a certain basis. These include symmetric matrix eigendecomposition, certain tensor decompositions, Independent Component Analysis (ICA), spectral clustering and Gaussian mixture learning. Each of these problems reduces to an instance of our general model, which we call a "Basis Encoding Function" (BEF). We show that learning a basis within this model can then be provably and efficiently achieved using a first order iteration algorithm (gradient iteration). Our algorithm goes beyond tensor methods while generalizing a number of existing algorithms---e.g., the power method for symmetric matrices, the tensor power iteration for orthogonal decomposable tensors, and cumulant-based FastICA---all within a broader function-based dynamical systems framework. Our framework also unifies the unusual phenomenon observed in these domains that they can be solved using efficient non-convex optimization. Specifically, we describe a class of BEFs such that their local maxima on the unit sphere are in one-to-one correspondence with the basis elements. This description relies on a certain "hidden convexity" property of these functions. We provide a complete theoretical analysis of the gradient iteration even when the BEF is perturbed. We show convergence and complexity bounds polynomial in dimension and other relevant parameters, such as perturbation size. Our perturbation results can be considered as a non-linear version of the classical Davis-Kahan theorem for perturbations of eigenvectors of symmetric matrices. In addition we show that our algorithm exhibits fast (superlinear) convergence and relate the speed of convergence to the properties of the BEF. Moreover, the gradient iteration algorithm can be easily and efficiently implemented in practice. Finally we apply our framework by providing the first provable algorithm for recovery in a general perturbed ICA model. |
14:30 | Multi-scale exploration of convex functions and bandit convex optimization (Best Paper Award) SPEAKER: Sébastien Bubeck ABSTRACT. 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 $\tilde{O}(\mathrm{poly}(n) \sqrt{T})$, 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. |
14:50 | Delay and Cooperation in Nonstochastic Bandits SPEAKER: unknown ABSTRACT. We study networks of communicating learning agents that cooperate to solve a common nonstochastic bandit problem. Agents use an underlying communication network to get messages about actions selected by other agents, and drop messages that took more than $d$ hops to arrive, where $d$ is a delay parameter. We introduce \textsc{Exp3-Coop}, a cooperative version of the {\sc Exp3} algorithm and prove that with $K$ actions and $N$ agents the average per-agent regret after $T$ rounds is at most of order $\sqrt{\bigl(d+1 + \tfrac{K}{N}\alpha_{\le d}\bigr)(T\ln K)}$, where $\alpha_{\le d}$ is the independence number of the $d$-th power of the connected communication graph $G$. We then show that for any connected graph, for $d=\sqrt{K}$ the regret bound is $K^{1/4}\sqrt{T}$, strictly better than the minimax regret $\sqrt{KT}$ for noncooperating agents. More informed choices of $d$ lead to bounds which are arbitrarily close to the full information minimax regret $\sqrt{T\ln K}$ when $G$ is dense. When $G$ has sparse components, we show that a variant of \textsc{Exp3-Coop}, allowing agents to choose their parameters according to their centrality in $G$, strictly improves the regret. Finally, as a by-product of our analysis, we provide the first characterization of the minimax regret for bandit learning with delay. |
15:10 | Policy Error Bounds for Model-Based Reinforcement Learning with Factored Linear Models SPEAKER: unknown ABSTRACT. In this paper we study a model-based approach to calculating approximately optimal policies in Markovian Decision Processes. In particular, we derive novel bounds on the loss of using a policy derived from a factored linear model, a class of models which generalize virtually all previous models that come with strong computational guarantees. For the first time in the literature, we derive performance bounds for model-based techniques where the model inaccuracy is measured in weighted norms. Moreover, our bounds show a decreased sensitivity to the discount factor and, unlike similar bounds derived for other approaches, they are insensitive to measure mismatch. Similarly to previous works, our proofs are also based on contraction arguments, but with the main differences that we use carefully constructed norms building on Banach lattices, and the contraction property is only assumed for operators acting on ``compressed'' spaces, thus weakening previous assumptions, while strengthening previous results. |
15:30 | Reinforcement Learning of POMDPs using Spectral Methods SPEAKER: unknown ABSTRACT. We propose a new reinforcement learning algorithm for partially observable Markov decision processes (POMDP) based on spectral decomposition methods. While spectral methods have been previously employed for consistent learning of (passive) latent variable models such as hidden Markov models, POMDPs are more challenging since the learner interacts with the environment and possibly changes the future observations in the process. We devise a learning algorithm running through episodes, in each episode we employ spectral techniques to learn the POMDP parameters from a trajectory generated by a fixed policy. At the end of the episode, an optimization oracle returns the optimal memoryless planning policy which maximizes the expected reward based on the estimated POMDP model. We prove an order-optimal regret bound with respect to the optimal memoryless policy and efficient scaling with respect to the dimensionality of observation and action spaces. |