Training a robot butler to make the perfect omlette could require breaking a lot of eggs and throwing out many imperfect attempts in a real-life kitchen. That's why researchers have been rolling out virtual training grounds as a more efficient alternative to putting AI agents through costly and time-consuming experiments in the real world. Virtual environments could prove especially useful in training the most popular AI based on machine learning algorithms that often require thousands of trial-and-error runs to learn new skills. Companies such as Waymo have already built their own internal simulators with virtual roads and traffic intersections to train their AI to safely take the wheel of self-driving cars. But a new, open-source virtual training ground called AI2-THOR enables AI agents to learn how to interact with objects in familiar home settings such as kitchens and bedrooms.

"AI is the new electricity" If you're a machine learning and ai enthusiast, you definitely must know this guy. He is best known for his machine learning course on coursera which, for many, has been the first step in understanding artificial intelligence(read my blog about it here). Andrew has been teaching at stanford ever since he got his Phd in 2002. He founded and led the google brain team which is considered as one of the most progressive ML/AI research organisations in the world. He also founded the popular massive open online course (MOOC) site coursera, which now has over a thousand courses taught by ivy league professors.

Translating is difficult work, the more so the further two languages are from one another. But sign language is a unique case, and translating it uniquely difficult, because it is fundamentally different from spoken and written languages. All the same, SignAll has been working hard for years to make accurate, real-time machine translation of ASL a reality. One would think that with all the advances in AI and computer vision happening right now, a problem as interesting and beneficial to solve as this would be under siege by the best of the best. Even thinking about it from a cynical market-expansion point of view, an Echo or TV that understands sign language could attract millions of new (and very thankful) customers.

For Valentine's day, new KDnuggets cartoon looks at some problems Machine Learning can face in 2118. Female Robot: Well, if you have not learned by now what I want, I am not going to tell you! This cartoon was ably drawn by Jon Carter. Here are other KDnuggets Valentine's Day Cartoons Cartoon: Perfect Valentine's Dates Found With Data Analysis, 2017 Cartoon: Data Scientist gets 3 wishes for Valentine's Day, 2015 Cartoon: Data Scientist Valentine Day Prediction, 2014. Data Scientist Valentine's Day Adjustment, 2013 See also other recent KDnuggets Cartoons: Cartoon: AI at Home: How Far Can A Smart Device Go? Cartoon: AI and Technology Transforming Christmas?

Precision medicine and artificial intelligence (AI) are complicated by design: Both scientific fields rely on extreme specificity, complex equations, and forces that can't be seen. As both fields begin to alter the healthcare landscape, they could plant a number of legal landmines. Can algorithms or biomarkers be patented? Will centers be able to access the large data sets they need to perform accurate AI? What control over their data should patients have?

Stochastic convex optimization algorithms are the most popular way to train machine learning models on large-scale data. Scaling up the training process of these models is crucial in many applications, but the most popular algorithm, Stochastic Gradient Descent (SGD), is a serial algorithm that is surprisingly hard to parallelize. In this paper, we propose an efficient distributed stochastic optimization method based on adaptive step sizes and variance reduction techniques. We achieve a linear speedup in the number of machines, small memory footprint, and only a small number of synchronization rounds -- logarithmic in dataset size -- in which the computation nodes communicate with each other. Critically, our approach is a general reduction than parallelizes any serial SGD algorithm, allowing us to leverage the significant progress that has been made in designing adaptive SGD algorithms. We conclude by implementing our algorithm in the Spark distributed framework and exhibit dramatic performance gains on large-scale logistic regression problems.

In recent years, Convolutional Neural Networks (CNNs) have shown remarkable performance in many computer vision tasks such as object recognition and detection. However, complex training issues, such as "catastrophic forgetting" and hyper-parameter tuning, make incremental learning in CNNs a difficult challenge. In this paper, we propose a hierarchical deep neural network, with CNNs at multiple levels, and a corresponding training method for lifelong learning. The network grows in a tree-like manner to accommodate the new classes of data without losing the ability to identify the previously trained classes. The proposed network was tested on CIFAR-10 and CIFAR-100 datasets, and compared against the method of fine tuning specific layers of a conventional CNN. We obtained comparable accuracies and achieved 40% and 20% reduction in training effort in CIFAR-10 and CIFAR 100 respectively. The network was able to organize the incoming classes of data into feature-driven super-classes. Our model improves upon existing hierarchical CNN models by adding the capability of self-growth and also yields important observations on feature selective classification.

For many machine learning problem settings, particularly with structured inputs such as sequences or sets of objects, a distance measure between inputs can be specified more naturally than a feature representation. However, most standard machine models are designed for inputs with a vector feature representation. In this work, we consider the estimation of a function $f:\mathcal{X} \rightarrow \R$ based solely on a dissimilarity measure $d:\mathcal{X}\times\mathcal{X} \rightarrow \R$ between inputs. In particular, we propose a general framework to derive a family of \emph{positive definite kernels} from a given dissimilarity measure, which subsumes the widely-used \emph{representative-set method} as a special case, and relates to the well-known \emph{distance substitution kernel} in a limiting case. We show that functions in the corresponding Reproducing Kernel Hilbert Space (RKHS) are Lipschitz-continuous w.r.t. the given distance metric. We provide a tractable algorithm to estimate a function from this RKHS, and show that it enjoys better generalizability than Nearest-Neighbor estimates. Our approach draws from the literature of Random Features, but instead of deriving feature maps from an existing kernel, we construct novel kernels from a random feature map, that we specify given the distance measure. We conduct classification experiments with such disparate domains as strings, time series, and sets of vectors, where our proposed framework compares favorably to existing distance-based learning methods such as $k$-nearest-neighbors, distance-substitution kernels, pseudo-Euclidean embedding, and the representative-set method.

I have a thoroughly delightful chance today to interview Vinod Khosla, who many years ago started Khosla Ventures, one of the most successful venture capital firms in the world. Vinod, I think you started out in engineering in India at one of the most prestigious institutes of technology, then you went to Carnegie Mellon, and then Stanford. But engineering wasn't where you landed long-term, because at some point you started Sun Microsystems, right? When I started out, there wasn't a thing called computer science. That tells you how old I am. In those days, you could pursue any new area if you created your own program because there were not a lot of programs. At the Indian Institute of Technology, we started the computer science program, and then I moved to biomedical engineering. I went on to get a master's degree in biomedical engineering at Carnegie Mellon. That was also a tiny program in the basement.

MIT researchers have developed a chip designed to speed up the hard work of running neural networks, while also reducing the power consumed when doing so dramatically – by up to 95 percent, in fact. The basic concept involves simplifying the chip design so that shuttling of data between different processors on the same chip is taken out of the equation. The big advantage of this new method, developed by a team lead by MIT graduate student Avishek Biswas, is that it could potentially be used to run neural networks on smartphones, household devices and other portable gadgets, rather than requiring servers drawing constant power from the grid. Because it means that phones of the future using this chip could do things like advanced speech and face recognition using neural nets and deep learning locally, rather than requiring on more crude, rule-based algorithms, or routing information to the cloud and back to interpret results. Computing'at the edge,' as its called, or at the site of sensors actually gathering the data, is increasingly something companies are pursuing and implementing, so this new chip design method could have a big impact on that growing opportunity should it become commercialized.