Nvidia cofounder Chris Malachowsky is eating a sausage omelet and sipping burnt coffee in a Denny's off the Berryessa overpass in San Jose. It was in this same dingy diner in April 1993 that three young electrical engineers--Malachowsky, Curtis Priem and Nvidia's current CEO, Jen-Hsun Huang--started a company devoted to making specialized chips that would generate faster and more realistic graphics for video games. East San Jose was a rough part of town back then--the front of the restaurant was pocked with bullet holes from people shooting at parked cop cars--and no one could have guessed that the three men drinking endless cups of coffee were laying the foundation for a company that would define computing in the early 21st century in the same way that Intel did in the 1990s. "There was no market in 1993, but we saw a wave coming," Malachowsky says. "There's a California surfing competition that happens in a five-month window every year. When they see some type of wave phenomenon or storm in Japan, they tell all the surfers to show up in California, because there's going to be a wave in two days. We were at the beginning."

Many activation functions have been proposed in the past, but selecting an adequate one requires trial and error. We propose a new methodology of designing activation functions within a neural network at each layer. We call this technique an "activation ensemble" because it allows the use of multiple activation functions at each layer. This is done by introducing additional variables, $\alpha$, at each activation layer of a network to allow for multiple activation functions to be active at each neuron. By design, activations with larger $\alpha$ values at a neuron is equivalent to having the largest magnitude. Hence, those higher magnitude activations are "chosen" by the network. We implement the activation ensembles on a variety of datasets using an array of Feed Forward and Convolutional Neural Networks. By using the activation ensemble, we achieve superior results compared to traditional techniques. In addition, because of the flexibility of this methodology, we more deeply explore activation functions and the features that they capture.

Human-robot interaction (HRI) is an emerging field of research with the aim to integrate robots into human social environments. One of the biggest challenges in the development of social robots is to understand human social norms [1]. It is therefore essential for social robots to possess deep models of social cognition, and be able to learn and adapt in accordance with their shared experiences with human partners. Most of the social robots to date are either preprogrammed, or are controlled by teleoperation or semiautonomous teleoperation [2], and do not possess the ability to learn and update themselves. Designing an adaptable and autonomous sociable robot is particularly challenging, as the robot needs to correctly interpret human behaviors as well as respond appropriately to them.

Although deep learning models have proven effective at solving problems in natural language processing, the mechanism by which they come to their conclusions is often unclear. As a result, these models are generally treated as black boxes, yielding no insight of the underlying learned patterns. In this paper we consider Long Short Term Memory networks (LSTMs) and demonstrate a new approach for tracking the importance of a given input to the LSTM for a given output. By identifying consistently important patterns of words, we are able to distill state of the art LSTMs on sentiment analysis and question answering into a set of representative phrases. This representation is then quantitatively validated by using the extracted phrases to construct a simple, rule-based classifier which approximates the output of the LSTM.

This time, Google's latest machine learning system, DeepMind, has learned to respond to stress with extreme aggression. I dunno about you, but that sounds like we just gave computers a fight or flight response. You may recall DeepMind as the computer that bested human Go players for the first time last years. Now, researchers have been using it to explore the limits of game theory -- a field of psychology that analyzes how people respond to cooperative and competitive opportunities. The team found that when DeepMind suspects that it's about to lose, it will switch to "highly aggressive" tactics to either win or maximize damage to its opponents.

Today's healthcare system was not built for a seamless integration of rapidly emerging technologies, such as machine learning innovations. Health data is largely inaccessible and not standardized, making it challenging to work with in machine learning systems. On top of this, deep learning techniques aren't well suited for many healthcare problems due to the difficulty in interpreting their decisions.

Alphabet Jigsaw (formerly Google Ideas) officially unveiled this morning their new tool for fighting toxic speech online, appropriately called Perspective. Powered by a deep learning model trained on more than 17 million manually reviewed reader comments provided by the New York Times, the model assigns a score to a given passage of text, rating it on a scale from 0 to 100% similar to statements that human reviewers have previously rated as "toxic." What makes this new approach from Google so different than past approaches is that it largely focuses on language rather than ideas: for the most part you can express your thoughts freely and without fear of censorship as long as you express them clinically and clearly, while if you resort to emotional diatribes and name calling, regardless of what you talk about, you will be flagged. What does this tell us about the future of toxic speech online and the notion of machines guiding humans to a more "perfect" humanity? One of the great challenges in filtering out "toxic" speech online is first defining what precisely counts as "toxic" and then determining how to remove such speech without infringing on people's ability to freely express their ideas.

MOOC Video Collection 777 views Decision Theory: Utility Functions - Stanford University - Duration: 18:16.

With huge strides in AI--from advances in the driverless vehicle realm, to mastering games such as poker and Go, to automating customer service interactions--this advanced technology is poised to revolutionize businesses. But the terms AI, machine learning, and deep learning are often used haphazardly and interchangeably, when there are key differences between each type of technology. Here's a guide to the differences between these three tools to help you master machine intelligence. AI is the broadest way to think about advanced, computer intelligence. In 1956 at the Dartmouth Artificial Intelligence Conference, the technology was described as such: "Every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."

When we asked dozens of venture capitalists where they see the most potential for applied artificial intelligence, they unanimously agreed on healthcare. Technology has already been used to incrementally improve patient medical records, care delivery, diagnostic accuracy, and drug development, but with A.I. we could achieve exponential breakthroughs. Deep learning first caught the media's attention when a team from the lab of Geoffrey Hinton at the University of Toronto won a Merck drug discovery competition despite having no experience with molecular biology and pharmaceutical development. Recently, a multidisciplinary research team at Stanford's School of Medicine comprised of pathologists, biomedical engineers, geneticists, and computer scientists developed deep learning algorithms that diagnose lung cancer more accurately than human pathologists. The ultimate dream in healthcare is to eradicate disease entirely.