SEATTLE--(BUSINESS WIRE)--Dato, creator of the popular machine learning platform GraphLab Create (GLC), announced today tools to give scientists, developers and users confidence in machine learning models and predictions. Dato is the first machine learning company to address the industry need for confidence in models and predictions. "Demand for machine learning has spread to large enterprise organizations," said Carlos Guestrin, Dato CEO and Amazon Professor of Machine Learning at University of Washington. "We have more than 80 commercial customers. The need for trust in models and predictions is an indicator of market adoption among established companies."

Many articles have been written about the top machine learning algorithms: click here and here for instance. Most of them seem to define top as oldest, and thus most used, ignoring modern, efficient algorithms fit for big data, such as indexation, attribution modeling, collaborative filtering, or recommendation engines used by companies such as Amazon, Google, or Facebook. I received this morning and advertisement for a (self-published) book called Master Machine Learning Algorithms, and I could not resist to post the author's list of top 10 machine learning algorithms:: Some of these techniques such as Naive Bayes (variables are almost never uncorrelated), Linear Discriminant Analysis (clusters are almost never separated by hyperplanes), or Linear Regression (numerous model assumptions - including linearity - are almost always violated in real data) have been so abused that I would hesitate teaching them. This is not a criticism of the book; most textbooks mention pretty much the same algorithms, and in this case, even skipping all graph-related algorithms. Even k Nearest Neighbors have modern, fast implementations not covered in traditional books - we are indeed working on this topic and expect to have an article published shortly about it.

If two people were to look at a common object, they may see entirely different things. According to a new study from Johns Hopkins University, a person's familiarity with an object - in particular with letters of the alphabet - will influence the features they notice. By studying the varying ways people perceive an alphabet, the researchers found that expertise helps to sort out the features that don't matter, leaving novices to view letters as more complex. While novices were found to point out differences more quickly, experts of the language were more accurate in their selections. As letters became more complex, decision time slowed for novices.

Latency to end-users and regulatory requirements push large companies to build data centers all around the world. The resulting data is "born" geographically distributed. On the other hand, many machine learning applications require a global view of such data in order to achieve the best results. These types of applications form a new class of learning problems, which we call Geo-Distributed Machine Learning (GDML). Such applications need to cope with: 1) scarce and expensive cross-data center bandwidth, and 2) growing privacy concerns that are pushing for stricter data sovereignty regulations. Current solutions to learning from geo-distributed data sources revolve around the idea of first centralizing the data in one data center, and then training locally. As machine learning algorithms are communication-intensive, the cost of centralizing the data is thought to be offset by the lower cost of intra-data center communication during training. In this work, we show that the current centralized practice can be far from optimal, and propose a system for doing geo-distributed training. Furthermore, we argue that the geo-distributed approach is structurally more amenable to dealing with regulatory constraints, as raw data never leaves the source data center. Our empirical evaluation on three real datasets confirms the general validity of our approach, and shows that GDML is not only possible but also advisable in many scenarios.

Empirical analysis serves as an important complement to theoretical analysis for studying practical Bayesian optimization. Often empirical insights expose strengths and weaknesses inaccessible to theoretical analysis. We define two metrics for comparing the performance of Bayesian optimization methods and propose a ranking mechanism for summarizing performance within various genres or strata of test functions. These test functions serve to mimic the complexity of hyperparameter optimization problems, the most prominent application of Bayesian optimization, but with a closed form which allows for rapid evaluation and more predictable behavior. This offers a flexible and efficient way to investigate functions with specific properties of interest, such as oscillatory behavior or an optimum on the domain boundary.

Models with latent variables have been proposed and investigated for explaining, understanding, or classifying observed data. If a model is a generative model, observed data are modeled to be as if they were generated by latent variables through parameterized probability distributions. Popular criteria for learning generative models include likelihood or posterior probability, which both evaluate the probability of the given observed data or parameters. Another kind of criteria is mutual information. Mutual information has been used to learn nonlinear generative models [14] in which relationships between observed and latent variables are directly evaluated. It has also been used to learn linear encoding (recognition) models [2, 12]. The relationships between observed and latent variables have greater importance in more complex generative models, e.g., deep learning models [6, 9]. In the pre-training of deep belief networks (DBNs), one of the models or techniques of deep learning, posterior samples of latent variables in the lower layer are used as samples of observed variables in the next, higher layer. For successive layer learning to be possible, latent variables should possess properties that enable such learning.

We present an approach to learning multi-sense word embeddings relying both on monolingual and bilingual information. Our model consists of an encoder, which uses monolingual and bilingual context (i.e. a parallel sentence) to choose a sense for a given word, and a decoder which predicts context words based on the chosen sense. The two components are estimated jointly. We observe that the word representations induced from bilingual data outperform the monolingual counterparts across a range of evaluation tasks, even though crosslingual information is not available at test time.

We study the robustness of active learning (AL) algorithms against prior misspecification: whether an algorithm achieves similar performance using a perturbed prior as compared to using the true prior. In both the average and worst cases of the maximum coverage setting, we prove that all $\alpha$-approximate algorithms are robust (i.e., near $\alpha$-approximate) if the utility is Lipschitz continuous in the prior. We further show that robustness may not be achieved if the utility is non-Lipschitz. This suggests we should use a Lipschitz utility for AL if robustness is required. For the minimum cost setting, we can also obtain a robustness result for approximate AL algorithms. Our results imply that many commonly used AL algorithms are robust against perturbed priors. We then propose the use of a mixture prior to alleviate the problem of prior misspecification. We analyze the robustness of the uniform mixture prior and show experimentally that it performs reasonably well in practice.

Off-policy reinforcement learning has many applications including: learning from demonstration, learning multiple goal seeking policies in parallel, and representing predictive knowledge. Recently there has been an proliferation of new policy-evaluation algorithms that fill a longstanding algorithmic void in reinforcement learning: combining robustness to off-policy sampling, function approximation, linear complexity, and temporal difference (TD) updates. This paper contains two main contributions. First, we derive two new hybrid TD policy-evaluation algorithms, which fill a gap in this collection of algorithms. Second, we perform an empirical comparison to elicit which of these new linear TD methods should be preferred in different situations, and make concrete suggestions about practical use.

Modal regression estimates the local modes of the distribution of $Y$ given $X=x$, instead of the mean, as in the usual regression sense, and can hence reveal important structure missed by usual regression methods. We study a simple nonparametric method for modal regression, based on a kernel density estimate (KDE) of the joint distribution of $Y$ and $X$. We derive asymptotic error bounds for this method, and propose techniques for constructing confidence sets and prediction sets. The latter is used to select the smoothing bandwidth of the underlying KDE. The idea behind modal regression is connected to many others, such as mixture regression and density ridge estimation, and we discuss these ties as well.