Sparse matrix factorization is a popular tool to obtain interpretable data decompositions, which are also effective to perform data completion or denoising. Its applicability to large datasets has been addressed with online and randomized methods, that reduce the complexity in one of the matrix dimension, but not in both of them. In this paper, we tackle very large matrices in both dimensions. We propose a new factorization method that scales gracefully to terabyte-scale datasets. Those could not be processed by previous algorithms in a reasonable amount of time. We demonstrate the efficiency of our approach on massive functional Magnetic Resonance Imaging (fMRI) data, and on matrix completion problems for recommender systems, where we obtain significant speedups compared to state-of-the art coordinate descent methods. Matrix factorization is a flexible tool for uncovering latent factors in low-rank or sparse models. For instance, building on low-rank structure, it has proven very powerful for matrix completion, e.g. in recommender systems (Srebro et al., 2004; Candès & Recht, 2009). In signal processing and computer vision, matrix factorization with a sparse regularization is often called dictionary learning and has proven very effective for denoising and visual feature encoding (see Mairal, 2014, for a review).

We study the problem of off-policy value evaluation in reinforcement learning (RL), where one aims to estimate the value of a new policy based on data collected by a different policy. This problem is often a critical step when applying RL to real-world problems. Despite its importance, existing general methods either have uncontrolled bias or suffer high variance. In this work, we extend the doubly robust estimator for bandits to sequential decision-making problems, which gets the best of both worlds: it is guaranteed to be unbiased and can have a much lower variance than the popular importance sampling estimators. We demonstrate the estimator's accuracy in several benchmark problems, and illustrate its use as a subroutine in safe policy improvement. We also provide theoretical results on the inherent hardness of the problem, and show that our estimator can match the lower bound in certain scenarios.

Once known to be used exclusively in military domain, unmanned aerial vehicles (drones) have stepped up to become a part of new logistic method in commercial sector called "last-mile delivery". In this novel approach, small unmanned aerial vehicles (UAV), also known as drones, are deployed alongside with trucks to deliver goods to customers in order to improve the service quality or reduce the transportation cost. It gives rise to a new variant of the traveling salesman problem (TSP), of which we call TSP with drone (TSP-D). In this article, we consider a variant of TSP-D where the main objective is to minimize the total transportation cost. We also propose two heuristics: "Drone First, Truck Second" (DFTS) and "Truck First, Drone Second" (TFDS), to effectively solve the problem. The former constructs route for drone first while the latter constructs route for truck first. We solve a TSP to generate route for truck and propose a mixed integer programming (MIP) formulation with different profit functions to build route for drone. Numerical results obtained on many instances with different sizes and characteristics are presented. Recommendations on promising algorithm choices are also provided.

This course is the next logical step in my deep learning, data science, and machine learning series. I've done a lot of courses about deep learning, and I just released a course about unsupervised learning, where I talked about clustering and density estimation. So what do you get when you put these 2 together? In these course we'll start with some very basic stuff - principal components analysis (PCA), and a popular nonlinear dimensionality reduction technique known as t-SNE (t-distributed stochastic neighbor embedding). Next, we'll look at a special type of unsupervised neural network called the autoencoder.

The four founders, two of whom have resigned their professorships to devote their full attention to TwentyBN, met each other during studies at the University of Bielefeld in Germany. Each of the founders has over 15 years of experience in machine learning and the relatively young deep learning discipline. Prof. Dr. Roland Memisevic, Chief Scientist, received his doctorate in Toronto, studying with Geoffrey Hinton, one of the founding fathers of deep learning. Prior to co-founding Twenty Billion Neurons, Memisevic was a member of the faculty at the renowned Machine Learning Institute of the University of Montréal led by Yoshua Bengio. The institute counts Google, Facebook, and IBM amongst its most active donors.

"Banks, financial services, insurance companies as well as global enterprises in other data-intensive industries are using cognitive automation to solve operational problems," said Alex Lyashok, COO of WorkFusion. "Gartner's report highlights the opportunity to use smart machines to transform existing business processes. Our product helps operations teams apply smart automation in an easy, non-disruptive way. Not only does WorkFusion help deliver as much as a 60% reduction in common back-office processes, but customers are also able to improve customer experience by reducing cycle times and improving the quality of customer interactions." WorkFusion is the leading smart process automation solution for enterprise operations.

You've made up your mind to become a data scientist. You've taken every data science MooC, you've eaten a lifetime of pizza at machine learning meetups, you even attended a data science "academy." Data Science is not knowledge to be acquired but rather a skill that can be learned and improved through practice. The number one qualification employers look for when hiring a data science candidate is previous experience. Startup.ML is launching a fellowship to give aspiring data scientists the chance to hone their skills by building real machine learning applications for startups and established data science teams.

The way you drive is surprisingly unique. And in an era when automobiles have become data-harvesting, multi-ton mobile computers, the data collected by your car--or one you rent or borrow--can probably identify you based on that driving style after as little as a few minutes behind the wheel. In a study they plan to present at the Privacy Enhancing Technology Symposium in Germany this July, a group of researchers from the University of Washington and the University of California at San Diego found that they could "fingerprint" drivers based only on data they collected from internal computer network of the vehicle their test subjects were driving, what's known as a car's CAN bus. In fact, they found that the data collected from a car's brake pedal alone could let them correctly distinguish the correct driver out of 15 individuals about nine times out of ten, after just 15 minutes of driving. With 90 minutes driving data or monitoring more car components, they could pick out the correct driver fully 100 percent of the time. "With very limited amounts of driving data we can enable very powerful and accurate inferences about the driver's identity," says Miro Enev, a former University of Washington researcher who worked on the study before taking a job as a machine-learning engineer at Belkin.

I write about search A LOT--about how to nail search engine marketing (SEM), the impact mobile has on consumers' search and purchasing habits and how RankBrain and artificial intelligence (AI) are changing the search game. Why do I write about this so much? That's easy: if you're a marketer and don't care about what's happening with search, it's impossible to do your job. The conversations about search so far, though, have always had one thing in common: We've been talking about text. Think for a moment about all the images and visual assets circulating the web.

What do you think of when you hear the word "technology"? Do you think of jet planes and laboratory equipment and underwater farming? Or do you think of smartphones and machine-learning algorithms? When a grave-faced announcer on CNBC says "technology stocks are down today," we all know he means Facebook and Apple, not Boeing and Pfizer. To Thiel, this signals a deeper problem in the American economy, a shrinkage in our belief of what's possible, a pessimism about what is really likely to get better. Our definition of what technology is has narrowed, and he thinks that narrowing is no accident. "Technology gets defined as'that which is changing fast,'" he says. "If the other things are not defined as'technology,' we filter them out and we don't even look at them." He founded PayPal and Palantir, was one of the earliest investors in Facebook, and now sits atop a fortune estimated in the billions. We spoke in his sleek, floor-to-ceiling-windowed apartment overlooking Manhattan -- a palace built atop the riches of the IT revolution.