maximum likelihood

Reward Augmented Maximum Likelihood for Neural Structured Prediction

Neural Information Processing Systems

A key problem in structured output prediction is enabling direct optimization of the task reward function that matters for test evaluation. This paper presents a simple and computationally efficient method that incorporates task reward into maximum likelihood training. We establish a connection between maximum likelihood and regularized expected reward, showing that they are approximately equivalent in the vicinity of the optimal solution. Then we show how maximum likelihood can be generalized by optimizing the conditional probability of auxiliary outputs that are sampled proportional to their exponentiated scaled rewards. We apply this framework to optimize edit distance in the output space, by sampling from edited targets.

Maximum Likelihood Learning With Arbitrary Treewidth via Fast-Mixing Parameter Sets

Neural Information Processing Systems

Inference is typically intractable in high-treewidth undirected graphical models, making maximum likelihood learning a challenge. One way to overcome this is to restrict parameters to a tractable set, most typically the set of tree-structured parameters. This paper explores an alternative notion of a tractable set, namely a set of "fast-mixing parameters" where Markov chain Monte Carlo (MCMC) inference can be guaranteed to quickly converge to the stationary distribution. While it is common in practice to approximate the likelihood gradient using samples obtained from MCMC, such procedures lack theoretical guarantees. This paper proves that for any exponential family with bounded sufficient statistics, (not just graphical models) when parameters are constrained to a fast-mixing set, gradient descent with gradients approximated by sampling will approximate the maximum likelihood solution inside the set with high-probability.

A Polynomial Time Algorithm for Log-Concave Maximum Likelihood via Locally Exponential Families

Neural Information Processing Systems

We consider the problem of computing the maximum likelihood multivariate log-concave distribution for a set of points. Specifically, we present an algorithm which, given $n$ points in $\mathbb{R} d$ and an accuracy parameter $\eps 0$, runs in time $\poly(n,d,1/\eps),$ and returns a log-concave distribution which, with high probability, has the property that the likelihood of the $n$ points under the returned distribution is at most an additive $\eps$ less than the maximum likelihood that could be achieved via any log-concave distribution. This is the first computationally efficient (polynomial time) algorithm for this fundamental and practically important task. Our algorithm rests on a novel connection with exponential families: the maximum likelihood log-concave distribution belongs to a class of structured distributions which, while not an exponential family, locally'' possesses key properties of exponential families. This connection then allows the problem of computing the log-concave maximum likelihood distribution to be formulated as a convex optimization problem, and solved via an approximate first-order method.

The Likelihood Principle, the MVUE, Ghosts, Cakes and Elves


In my prior blog post, I wrote of a clever elf that could predict the outcome of a mathematically fair process roughly ninety percent of the time. Actually, it is ninety-three percent of the time and why it is ninety-three percent instead of ninety percent is also important. The purpose of the prior blog post was to illustrate the weakness of using the minimum variance unbiased estimator (MVUE) in applied finance. Nonetheless, that begs a more general question of when and why it should be used, or a Bayesian or Likelihood-based method should be applied. Fortunately, the prior blog post provides a way of looking at the problem. Fisher's Likelihood-based, Pearson and Neyman's Frequency-based and Laplace's method of inverse probability really are at odds with one another. Indeed, much of the literature of the mid-twentieth century had a polemical ring to it.

Representation Learning: A Statistical Perspective Machine Learning

Learning representations of data is an important problem in statistics and machine learning. While the origin of learning representations can be traced back to factor analysis and multidimensional scaling in statistics, it has become a central theme in deep learning with important applications in computer vision and computational neuroscience. In this article, we review recent advances in learning representations from a statistical perspective. In particular, we review the following two themes: (a) unsupervised learning of vector representations and (b) learning of both vector and matrix representations.

Balancing Act in Datasets of a Machine Learning algorithm


When dealing with imbalanced classes, we may need to do some extra work and planning to make sure that our algorithms give us useful results. In this blog, I examine just two classification techniques to illustrate the issue, but you should know that the problem generalizes. For good reason, supervised classification algorithms -- which use labeled data -- take class distributions into account. However, when we're trying to detect classes that are important, but rare compared to the alternatives, it can be difficult to develop a model that catches them. Here, after diving into the problem with some examples, I outline a few of the tried and true techniques for solving it.

How Bayes' Theorem is Applied in Machine Learning - KDnuggets


In the previous post we saw what Bayes' Theorem is, and went through an easy, intuitive example of how it works. You can find this post here. If you don't know what Bayes' Theorem is, and you have not had the pleasure to read it yet, I recommend you do, as it will make understanding this present article a lot easier. In this post, we will see the uses of this theorem in Machine Learning. As mentioned in the previous post, Bayes' theorem tells use how to gradually update our knowledge on something as we get more evidence or that about that something.

A Gentle Introduction to Information Entropy


Information theory is a subfield of mathematics concerned with transmitting data across a noisy channel. A cornerstone of information theory is the idea of quantifying how much information there is in a message. More generally, this can be used to quantify the information in an event and a random variable, called entropy, and is calculated using probability. Calculating information and entropy is a useful tool in machine learning and is used as the basis for techniques such as feature selection, building decision trees, and, more generally, fitting classification models. As such, a machine learning practitioner requires a strong understanding and intuition for information and entropy.

Logistic Regression


A member of the generalized linear model (GLM) family and similar to linear regression in many ways, logistic regression (despite the confusing name) is used for classification problems with two possible outcomes. Logistic regression is handy for classification problems since it fits an S shaped logistic (or Sigmoid) function to the data, squishing the linear equation to an output range of 0–1. This convenient range allows logistic regression to model the probabilities of a data point belonging to a particular class, typically with the decision point at the probability of .5. So, what does that look like in math? How does the sigmoid function squish the linear equation?

Wasserstein Neural Processes Machine Learning

Neural Processes (NPs) are a class of models that learn a mapping from a context set of input-output pairs to a distribution over functions. They are traditionally trained using maximum likelihood with a KL divergence regularization term. We show that there are desirable classes of problems where NPs, with this loss, fail to learn any reasonable distribution. We also show that this drawback is solved by using approximations of Wasserstein distance which calculates optimal transport distances even for distributions of disjoint support. We give experimental justification for our method and demonstrate performance. These Wasserstein Neural Processes (WNPs) maintain all of the benefits of traditional NPs while being able to approximate a new class of function mappings.