# New paper: Redundancy of Exchangeable Estimators

More like an old paper… this (finally) a journal version of some older work that we did on analyzing Bayesian nonparametric estimators form an information-theoretic (redundancy) perspective.

Exchangeable random partition processes are the basis for Bayesian approaches to statistical inference in large alphabet settings. On the other hand, the notion of the pattern of a sequence provides an information-theoretic framework for data compression in large alphabet scenarios. Because data compression and parameter estimation are intimately related, we study the redundancy of Bayes estimators coming from Poisson-Dirichlet priors (or “Chinese restaurant processes”) and the Pitman-Yor prior. This provides an understanding of these estimators in the setting of unknown discrete alphabets from the perspective of universal compression. In particular, we identify relations between alphabet sizes and sample sizes where the redundancy is small, thereby characterizing useful regimes for these estimators.

In the large alphabet setting, one thing we might be interested in is sequential prediction: I observe a sequence of butterfly species and want to predict whether the next butterfly I collect will be new or one that I have seen before. One simple way to do this prediction is to put a prior on the set of all distributions on infinite supports and do inference on that model given the data. This corresponds to the so-called Chinese Restaurant Process (CRP) approach to the problem. The information-theoretic view is that sequential prediction is equivalent to compression: the estimator is assigning a probability $q(x^n)$ to the sequence $x^n$ seen so far. An estimator is good if for any distribution $p$, if $x^n$ is drawn i.i.d. according to $p$, then the divergence between $p(x^n)$ and $q(x^n)$ is “small.” The goal of this work is to understand when CRP estimators are good in this sense.

This sort of falls in with the “frequentist analysis of Bayesian procedures” thing which some people work on.

# ICML 2014: a few more papers

After a long stint of proposal writing, I figured I should catch up on some old languishing posts. So here’s a few quick notes on the remainder of ICML 2014.

• Fast Stochastic Alternating Direction Method of Multipliers (Wenliang Zhong; James Kwok): Most of the talks in the Optimization II session were on ADMM or stochastic optimization, or both. This was int he last category. ADMM can have rather high-complexity update rules, especially on large, complex problems, so the goal is to lower the complexity of the update step by making it stochastic. The hard part seems to be controlling the step size.
• An Asynchronous Parallel Stochastic Coordinate Descent Algorithm (Ji Liu; Steve Wright; Christopher Re; Victor Bittorf; Srikrishna Sridhar): The full version of this paper is on ArXiV. The authors look at a multi-core lock-free stochastic coordinate descent method and characterize how many cores you need to get linear speedups — this depends on the convexity properties of the objective function.
• Communication-Efficient Distributed Optimization using an Approximate Newton-type Method (Ohad Shamir; Nati Srebro; Tong Zhang): This paper looked 1-shot “average at the end” schemes where you divide the data onto multiple machines, have them each train a linear predictor (for example) using stochastic optimization and then average the results. This is just averaging i.i.d. copies of some complicated random variable (the output of an optimization) so you would expect some variance reduction. This method has been studied by a few people int the last few years. While you do get variance reduction, the bias can still be bad. On the other extreme, communicating at every iteration essentially transmits the entire data set (or worse) over the network. They propose a new method for limiting communication by computing an approximate Newton step without approximating the full Hessian. It works pretty well.
• Lower Bounds for the Gibbs Sampler over Mixtures of Gaussians (Christopher Tosh; Sanjoy Dasgupta): This was a great talk about how MCMC can be really slow to converge. The model is a mixture of Gaussians with random weights (Dirichlet) and means (Gaussian I think). Since the posterior on the parameters is hard to compute, you might want to do Gibbs sampling. They use conductance methods to get a lower bound on the mixing time of the chain. The tricky part is that the cluster labels are permutation invariant — I don’t care if you label clusters (1,2) versus (2,1), so they need to construct some equivalence classes. They also have further results on what happens when the number of clusters is misspecified. I really liked this talk because MCMC always seems like black magic to me (and I even used it in a paper!)
• (Near) Dimension Independent Risk Bounds for Differentially Private Learning (Prateek Jain; Abhradeep Guha Thakurta): Abhradeep presented a really nice paper with a tighter analysis of output and objective perturbation methods for differentially private ERM, along with a new algorithm for risk minimization on the simplex. Abhradeep really only talked about the first part. If you focus on scalar regret, they show that essentially the error comes from taking the inner product of a noise vector with a data vector. If the noise is Gaussian then the noise level is dimension-independent for bounded data. This shows that taking $(\epsilon,\delta)$-differential privacy yield better sample complexity results than $(\epsilon,)$-differential privacy. This feels similar in flavor to a recent preprint on ArXiV by Beimel, Nissim, and Stemmer.
• Near-Optimally Teaching the Crowd to Classify (Adish Singla; Ilija Bogunovic; Gabor Bartok; Amin Karbasi; Andreas Krause): This was one of those talks where I would have to go back to look at the paper a bit more. The idea is that you want to train annotators to do better in a crowd system like Mechanical Turk — which examples should you give them to improve their performance? They model the learners as doing some multiplicative weights update. Under that model, the teacher has to optimize to pick a batch of examples to give to the learner. This is hard, so they use a submodular surrogate function and optimize over that.
• Discrete Chebyshev Classifiers (Elad Eban; Elad Mezuman; Amir Globerson): This was an award-winner. The setup is that you have categorical (not numerical) features on $n$ variables and you want to do some classification. They consider taking pairwise inputs and compute for each tuple $(x_i, x_j, y)$ a marginal $\mu_{ij}(x_i, x_j, y)$. If you want to create a rule $f: \mathcal{X} \to \mathcal{Y}$ for classification, you might want to pick one that has best worst-case performance. One approach is to take the one which has best worst-case performance over all joint distributions on all variables that agree with the empirical marginals. This optimization looks hard because of the exponential number of variables, but they in fact show via convex duality and LP relaxations that it can be solved efficiently. To which I say: wow! More details are in the paper, but the proofs seem to be waiting for a journal version.

# LaTeX typeface fashions of the early ’90s

While trying to show a student a generic example of a paper’s structure, I came across this gem:

I feel like I am reading a MacWrite document while wearing a flannel shirt.