Reporting top-1 one-crop accuracy versus amount of operations required for a single forward pass in multiple popular neural network architectures. At the time GPU offered a much larger number of cores than CPUs, and allowed 10x faster training time, which in turn allowed to use larger datasets and also bigger images. Christian thought a lot about ways to reduce the computational burden of deep neural nets while obtaining state-of-art performance (on ImageNet, for example). Inspired by NiN, the bottleneck layer of Inception was reducing the number of features, and thus operations, at each layer, so the inference time could be kept low.

Nvidia, the California-based graphics chip manufacturer and technology company, was officially crowned by the Massachusetts Institute of Technology (MIT) as the smartest company in the world for its innovation in fields such as AI, deep learning and AI. The list of "smart" companies produced by MIT included some of the most innovative companies in the technology and manufacturing sectors such as SpaceX, Tesla and Amazon but according to MIT Technology Review editor David Rotman, Nvidia's explosive growth in the AI and Blockchain markets gave the firm an edge over other companies. "The list is our best guess as to which firms will be the dominant companies of the future. Amazon and Facebook and Google are on it, but so are plenty of newcomers," wrote Rotman. Jen-Hsun Huang, CEO of Nvidia, recently revealed the company's latest deep learning-specific chip the Tesla P100, which the company spent over $2 bln in research and development to manufacturing.

"Yes, we don't know how an AI builds its model, but neither do I know how I catch a ball." This was Tim O'Reilly's tweeted response to an article of mine (https://goo.gl/5vqNfg) It's a good point –O'Reilly is the founder of O'Reilly Media and is an exceptionally smart person--both in general and when considering Deep Learning. Knowledge is always the surface of a deep pool of ignorance. This is for at least two reasons.

Artificial intelligence is already being put to use in the NHS, with Google's AI firm DeepMind providing technology to help monitor patients. And a new study suggests that Google could soon be meeting with Genomic England - a company set up by the Department of Health to sequence 100,000 genomes – to discuss whether DeepMind could get involved. In an article for The Conversation, Edward Hockings a researcher at the University of the West of Scotland, explains the risks of letting a private company gain access to sensitive genetic data. In Google's case, he says, it could allow them to target users with personalised advertising based on their preferences and health risks. It could also create profiles of people based on their DNA data, which may provide details such as their risk of becoming a criminal.

Most current AI systems can be easily fooled, which is a problem that affects almost all machine learning techniques. Deep neural networks have millions of parameters and to understand why the network provides good or bad results becomes impossible.

Talos was a giant bronze warrior created to guard the island of Crete from pirates and invaders. He circled the island three times daily, and his menacing appearance encouraged would-be pirates to seek treasure elsewhere. A scarecrow, made only to project the image of a warrior. The faithful however, believed that the craftsmen had imbued creations such as Talos with very real minds, capable of emotion, thought and wisdom. This was of course, false. Talos was simply the latest manifestation of a dream that has consumed the minds of intellectuals for almost all of human history: the desire to create life-like, intelligent beings such as ourselves.

Scaling up systems for resource-intensive machine learning tasks demands convenient methods to manage computations distributed across multiple servers. Come and learn about both the processes underlying new Deep Learning techniques that have been applied to piloting drones and driving autonomous vehicles as well as the Docker containerization tools used to train these systems at scale.

Welcome to the artificial intelligence era. Unless you have been hiding under a rock for the last decade, you will have come across some form of artificial intelligence tools or solutions in your life. If you are anything like me, you have been excited to welcome the innovations brought on by artificial intelligence. But when it comes to understanding the landscape of the artificial intelligence ecosystem, it can truly be confusing. This ecosystem includes terms such as general artificial intelligence, artificial narrow intelligence, machine learning, deep learning, and so many others.

Compressing neural nets is an active research problem, given the large size of state-of-the-art nets for tasks such as object recognition, and the computational limits imposed by mobile devices. We give a general formulation of model compression as constrained optimization. This includes many types of compression: quantization, low-rank decomposition, pruning, lossless compression and others. Then, we give a general algorithm to optimize this nonconvex problem based on the augmented Lagrangian and alternating optimization. This results in a "learning-compression" algorithm, which alternates a learning step of the uncompressed model, independent of the compression type, with a compression step of the model parameters, independent of the learning task. This simple, efficient algorithm is guaranteed to find the best compressed model for the task in a local sense under standard assumptions. We present separately in several companion papers the development of this general framework into specific algorithms for model compression based on quantization, pruning and other variations, including experimental results on compressing neural nets and other models.

We propose a generalization of neural network sequence models. Instead of predicting one symbol at a time, our multi-scale model makes predictions over multiple, potentially overlapping multi-symbol tokens. A variation of the byte-pair encoding (BPE) compression algorithm is used to learn the dictionary of tokens that the model is trained with. When applied to language modelling, our model has the flexibility of character-level models while maintaining many of the performance benefits of word-level models. Our experiments show that this model performs better than a regular LSTM on language modeling tasks, especially for smaller models.