Virtually everywhere you look, Bay Area tech businesses are running into walls. Smartphones were revolutionary and lucrative, but the U.S. market is saturated, and Apple's iPhone sales have fallen for three quarters. The "app economy" has matured, with more people using existing apps than downloading new ones. And Facebook, which has filled users' news feeds with so many ads it can barely add more, is predicting its revenue growth will slump next year. Silicon Valley needs its next big thing, a focus for the concentrated brain power and innovation infrastructure that have made this region the world leader in transformative technology.

As Google learned, predicting the spread of influenza, even with mountains of data, is notoriously difficult. Nonetheless, bioinformatician and R user Shirin Glander has created a two-part tutorial about predicting flu deaths with R (part 2 here). The analysis is based on just 136 cases of influenza A H7N9 in China in 2013 (data provided in the outbreaks package) so the intent was not to create a generally predictive model, but by providing all of the R code and graphics Shirin has created a useful example of real-word predictive modeling with R. The tutorial covers loading and cleaning the data (including a nice example of using the mice package to impute missing values) and begins with some exploratory data visualizations. I was particularly impressed by the use of density charts (using the stat_density2d ggplot2 aesthetic) to highlight differences in the scatterplots of flu cases ending in death and recovery. Decision trees (implemented using rpart and visualized using fancyRpartPlot from the rattle package) Random Forests (using caret's "rf" training method) Elastic-Net Regularized Generalized Linear Models (using caret's "glmnet" training method) K-nearest neighbors clustering (using caret's "kknn" training method) Penalized Discriminant Analysis (using caret's "pda" training method) and in Part 2, Extreme gradient boosting using the xgboost package and various preprocessing techniques from the caret package Due to the limited data size, there's not too much difference between the models: in each case, 13-15 of the 23 cases were classified correctly.

We recently launched LinkedIn Learning, an online learning platform that enables students and professionals to take courses and learn the skills required to meet their career goals. As part of this platform, we provide personalized course recommendations. A/B testing indicates that we have 58% higher engagement rate when we provide personalized recommendations compared to generic or randomized recommendations. It's important to call out that these personalized recommendations are made possible by the robust, highly-structured knowledge base of member-skill-job connections that we have assembled at LinkedIn. For more information about this foundational work that enables machine learning and relevance at LinkedIn, please refer to Building the LinkedIn Knowledge Graph by Qi He, Bee-Chung Chen, and Deepak Agarwal.

Our smartphone currently represents the most expensive area to be purchased per squared centimeter (even more expensive than the square meters price of houses in Beverly Hills), and it is not hard to envision that having a bot as unique interfaces will make this area worth almost zero. None of these would be possible though without heavily investing in speech recognition research. Deep Reinforcement Learning (DFL) has been the boss in town for the past few years and it has been fed by human feedbacks. However, I personally believe that soon we will move toward a B2B (bot-to-bot) training for a very simple reason: the reward structure. Humans spend time training their bots if they are enough compensated for their effort.

Artificial Intelligence (AI) is an important and evolving concept that is having significant impact within the Customer Experience industry -- and it's a topic that is being talked about on a seemingly daily basis at this point. But is AI really ready for primetime in customer care? I spoke with Michael Johnston, Lead Inventive Scientist at Interactions, about frequently asked questions about AI and Machine Learning as they apply to customer care. How do you define Artificial Intelligence and Machine Learning? Artificial Intelligence refers to the capability of a machine to imitate intelligent human behavior.

We study the complexity of the problem of searching for a set of patterns that separate two given sets of strings. This problem has applications in a wide variety of areas, most notably in data mining, computational biology, and in understanding the complexity of genetic algorithms. We show that the basic problem of finding a small set of patterns that match one set of strings but do not match any string in a second set is difficult (NP-complete, W[2]-hard when parameterized by the size of the pattern set, and APX-hard). We then perform a detailed parameterized analysis of the problem, separating tractable and intractable variants. In particular we show that parameterizing by the size of pattern set and the number of strings, and the size of the alphabet and the number of strings give FPT results, amongst others.

Philosophers and psychologists have always been interested in the deep nature of human concepts, how they are formed, how they combine to create more complex conceptual structures, as expressed by sentences and texts, and how meaning is created in these processes. Unveiling aspects of these mysteries is bound to have a massive impact on a variety of domains, from knowledge representation to natural language processing, machine learning and artificial intelligence. The original idea of a concept as a'container of objects', called'instantiations', which can be traced back to Aristotle, was challenged by the first cognitive tests by Eleanor Rosch, which revealed that concepts exhibit aspects, like'context-dependence', 'vagueness' and'graded typicality', that prevent a too naïve definition of a concept as a'set of defining properties that are either possessed or not possessed by individual exemplars' [1, 2]. More, these tests infused the suspicion that concepts do not combine by following the algebraic rules of classical logic. A first attempt to preserve a set theoretical modeling came from the'fuzzy set approach': concepts would be represented by fuzzy sets, while their conjunction (disjunction) satisfies the'minimum (maximum) rule of fuzzy set conjunction (disjunction)' [3]. However, also this approach was confuted by a whole set of experiments by cognitive psychologists, including Osherson and Smith, who identified the'Guppy effect' (or'Pet-Fish problem') in typicality judgments [4], James Hampton, who discovered'overextension' and'underextension' effects in membership judgments [5, 6], and Alxatib and Pelletier, who detected'borderline contradictions' in simple propositions of the form "John is tall and John is not tall" [7]. More recently, some of us proved that these data violate Kolmogorov's axioms of classical probability theory [8], thus revealing that classical structures,

Representing a dialog policy as a recurrent neural network (RNN) is attractive because it handles partial observability, infers a latent representation of state, and can be optimized with supervised learning (SL) or reinforcement learning (RL). For RL, a policy gradient approach is natural, but is sample inefficient. In this paper, we present 3 methods for reducing the number of dialogs required to optimize an RNN-based dialog policy with RL. The key idea is to maintain a second RNN which predicts the value of the current policy, and to apply experience replay to both networks. On two tasks, these methods reduce the number of dialogs/episodes required by about a third, vs. standard policy gradient methods.

We propose a first data-driven tuning procedure for divide-and-conquer kernel ridge regression (Zhang et al., 2015). While the proposed criterion is computationally scalable for massive data sets, it is also shown to be asymptotically optimal under mild conditions. The effectiveness of our method is illustrated by extensive simulations and an application to Million Song Dataset. Some key words:Distributed GCV, divide-and-conquer, kernel ridge regression, optimal tuning.

Non-convex and non-smooth optimization plays an important role in machine learning. Proximal gradient method is one of the most important methods for solving the nonconvex and non-smooth problems, where a proximal operator need to be solved exactly for each step. However, in a lot of problems the proximal operator does not have an analytic solution, or is expensive to obtain an exact solution. In this paper, we propose inexact proximal gradient methods (not only a basic inexact proximal gradient method (IPG), but also a Nesterov's accelerated inexact proximal gradient method (AIPG)) for non-convex and non-smooth optimization, which tolerate an error in the calculation of the proximal operator. Theoretical analysis shows that IPG and AIPG have the same convergence rates as in the error-free case, provided that the errors decrease at appropriate rates. Keywords: Non-convex optimization, non-smooth optimization, proximal gradient, inexact proximal operator, Nesterov's accelerated method