The most important subpackages provide implementations of boilerplate code that is relevant to machine-learning problems. These include computer vision, natural language processing, and speech processing. Other subpackages may be smaller and focus on more specific problems or even specific data sets.

SITTING IN AN office in San Francisco, Igor Barani calls up some medical scans on his screen. He is the chief executive of Enlitic, one of a host of startups applying deep learning to medicine, starting with the analysis of images such as X-rays and CT scans. It is an obvious use of the technology. Deep learning is renowned for its superhuman prowess at certain forms of image recognition; there are large sets of labelled training data to crunch; and there is tremendous potential to make health care more accurate and efficient. Dr Barani (who used to be an oncologist) points to some CT scans of a patient's lungs, taken from three different angles.

As buzzwords become ubiquitous they become easier to tune out. We've finely honed this defense mechanism, for good purpose. It's better to focus on what's in front of us than the flavor of the week. CRISPR might change our lives, but knowing how it works doesn't help you. VR could eat all media, but it's hardware requirements keep it many years away from common use.

A curated list of resources dedicated to recurrent neural networks (closely related to deep learning). Please feel free to pull requests, email Myungsub Choi (cms6539@gmail.com)

Every company is now a data company, capable of using machine learning in the cloud to deploy intelligent apps at scale, thanks to three machine learning trends: data flywheels, the algorithm economy, and cloud-hosted intelligence. That was the takeaway from the inaugural Machine Learning / Artificial Intelligence Summit, hosted by Madrona Venture Group* last month in Seattle, where more than 100 experts, researchers, and journalists converged to discuss the future of artificial intelligence, trends in machine learning, and how to build smarter applications. With hosted machine learning models, companies can now quickly analyze large, complex data, and deliver faster, more accurate insights without the high cost of deploying and maintaining machine learning systems. "Every successful new application built today will be an intelligent application," Soma Somasegar said, venture partner at Madrona Venture Group. "Intelligent building blocks and learning services will be the brains behind apps."

Google has released technical paper on AI safety produced in collaboration with researchers from Stanford University, University of California, Berkeley and OpenAI.

Specialized electronic circuits called graphic processing units, or GPUs, are at the heart of modern mobile phones, personal computers and gaming consoles. By combining multiple GPUs in concert, researchers can solve previously elusive image processing problems. For example, Google and Facebook have both developed extremely accurate facial recognition software using these new techniques. GPUs are also crucial to radiologists, because they can rapidly process large medical imaging datasets from CT, MRI, ultrasound and even conventional x-rays. Now some radiology groups and technology companies are combining multiple GPUs with artificial intelligence (AI) algorithms to help improve radiology care.

In a Monday blog post, the leadership of artificial intelligence (AI) research company OpenAI said that the group wants to modify'off-the-shelf' robots so they can perform common household tasks. "We're working to enable a physical robot (off-the-shelf; not manufactured by OpenAI) to perform basic housework," the group said in a blog post authored by Research Director Ilya Sutskever, Chief Technology Officer Greg Brockman, Sam Altman and Elon Musk. This futuristic target is second only to the primary goal laid out in the organization's blog post, which is to develop AI that could learn to improve its ability over time. Meeting such a goal would provide an underpinning for the perhaps more glamorous concept of robots that can clean your home, but the post goes onto say that domestic robots themselves would provide a solid foundation for approaching other problems in AI. "There are existing techniques for specific tasks, but we believe that learning algorithms can eventually be made reliable enough to create a general-purpose robot. More generally, robotics is a good testbed for many challenges in AI," the blog post reads.

We introduce structured prediction energy networks (SPENs), a flexible framework for structured prediction. A deep architecture is used to define an energy function of candidate labels, and then predictions are produced by using back-propagation to iteratively optimize the energy with respect to the labels. This deep architecture captures dependencies between labels that would lead to intractable graphical models, and performs structure learning by automatically learning discriminative features of the structured output. One natural application of our technique is multi-label classification, which traditionally has required strict prior assumptions about the interactions between labels to ensure tractable learning and prediction. We are able to apply SPENs to multi-label problems with substantially larger label sets than previous applications of structured prediction, while modeling high-order interactions using minimal structural assumptions. Overall, deep learning provides remarkable tools for learning features of the inputs to a prediction problem, and this work extends these techniques to learning features of structured outputs. Our experiments provide impressive performance on a variety of benchmark multi-label classification tasks, demonstrate that our technique can be used to provide interpretable structure learning, and illuminate fundamental trade-offs between feed-forward and iterative structured prediction.

We present NN-grams, a novel, hybrid language model integrating n-grams and neural networks (NN) for speech recognition. The model takes as input both word histories as well as n-gram counts. Thus, it combines the memorization capacity and scalability of an n-gram model with the generalization ability of neural networks. We report experiments where the model is trained on 26B words. NN-grams are efficient at run-time since they do not include an output soft-max layer. The model is trained using noise contrastive estimation (NCE), an approach that transforms the estimation problem of neural networks into one of binary classification between data samples and noise samples. We present results with noise samples derived from either an n-gram distribution or from speech recognition lattices. NN-grams outperforms an n-gram model on an Italian speech recognition dictation task.