We consider neural network training, in applications in which there are many possible classes, but at test-time, the task is to identify only whether the given example belongs to a specific class, which can be different in different applications of the classifier. For instance, this is the case in an image search engine. We consider the Single Logit Classification (SLC) task: training the network so that at test-time, it would be possible to accurately identify if the example belongs to a given class, based only on the output logit for this class. We propose a natural principle, the Principle of Logit Separation, as a guideline for choosing and designing losses suitable for the SLC. We show that the cross-entropy loss function is not aligned with the Principle of Logit Separation. In contrast, there are known loss functions, as well as novel batch loss functions that we propose, which are aligned with this principle. In total, we study seven loss functions. Our experiments show that indeed in almost all cases, losses that are aligned with Principle of Logit Separation obtain a 20%-35% relative performance improvement in the SLC task, compared to losses that are not aligned with it. We therefore conclude that the Principle of Logit Separation sheds light on an important property of the most common loss functions used by neural network classifiers.

Knowledge graphs enable a wide variety of applications, including question answering and information retrieval. Despite the great effort invested in their creation and maintenance, even the largest (e.g., Yago, DBPedia or Wikidata) remain incomplete. We introduce Relational Graph Convolutional Networks (R-GCNs) and apply them to two standard knowledge base completion tasks: Link prediction (recovery of missing facts, i.e. subject-predicate-object triples) and entity classification (recovery of missing entity attributes). R-GCNs are related to a recent class of neural networks operating on graphs, and are developed specifically to deal with the highly multi-relational data characteristic of realistic knowledge bases. We demonstrate the effectiveness of R-GCNs as a stand-alone model for entity classification. We further show that factorization models for link prediction such as DistMult can be significantly improved by enriching them with an encoder model to accumulate evidence over multiple inference steps in the relational graph, demonstrating a large improvement of 29.8% on FB15k-237 over a decoder-only baseline.

Hewlett Packard Enterprise (HPE) is tying artificial intelligence (AI) into a handful of new products and services, with a specific focus on enhancing deep learning. The new offerings include hardware, software, reference designs, and physical research locations. On the hardware and software front, HPE launched an integrated product that ties its Apollo 6500 hardware box with software from technology partner Bright Computing. The combination is designed to allow for deep learning application development using pre-configured software frameworks, libraries, automated software updates, and cluster management. The AI Research unit at Hewlett Packard Labs also unveiled a set of tools to help customers in selecting hardware and software environments for different deep learning tasks.

Are you promoting AI to consumers, clients or colleagues? If you're in marketing today there is a good chance that you are. But how do you sell artificial intelligence effectively given all the hype and hysteria that surrounds the technology? Here are ten evidence-based recommendations for how to communicate AI effectively from a new Syzygy study that captured people's feelings towards AI across the US, UK, and Germany (n 6000). With all the hype and hysteria around AI, people are suspicious and skeptical.

It has become common to joke about how robots are going to take our jobs, and rightfully so: Oxford University researchers estimate that 47% of all current US employment is at high risk to become automated over the next decade or so. But there is positive news: Of the 1.8 million jobs AI will get rid of, the emerging field will create 2.3 million by 2020, according to a recent report from Gartner. And a recent Capgemini report found that 83% of companies using AI say the technology is already adding jobs. A lot of that growth is coming from the technology itself. "We'll continue to see job growth in anything AI-related for the next five to 10 years, which is one of the things that will mitigate the oft-publicized inevitable job loss due to AI-led automation," said Brandon Purcell, an analyst at Forrester.

Because we know how to have fun, we were (litteraly) excited about spending days and nights on crunching data of European startups funded in 2016 to eventually get a real sense of what is AI and Data's real trend in Europe. Here are 4 facts that will make you change your views and you can discover the full report at the end. Numbers stagger: $774 million -- Yes 7–7 and 4, nearly 10% of the total 10 billion invested in European startups in 2016, has been fully dedicated to AI & data. Even more striking is the number of AI start-ups financed: 271, three times 2015 figures. This trend has been led by early stage investments which played a signifiant role: $215 million have been invested in 171 early stage startups (0 to $5m in funding).

Is there a global theory for the shapes of fish? But for most of the history of biology, it's not the kind of thing anyone would ever have asked. Stephen Wolfram is the creator of Mathematica, Wolfram Alpha and the Wolfram Language; the author of A New Kind of Science; and the founder and CEO of Wolfram Research. Sign up to get Backchannel's weekly newsletter, and follow us on Facebook, Twitter, and Instagram. And it's now 100 years since D'Arcy Thompson published the first edition of his magnum opus On Growth and Form--and tried to use ideas from mathematics and physics to discuss global questions of biological growth and form. Stretch one kind of fish, and it looks like another. Yes, without constraints on how you stretch. It's not quite clear what this is telling one, and I don't think it's much. But just to ask the question is interesting, and On Growth and Form is full of interesting questions--together with all manner of curious and interesting answers. D'Arcy Thompson was in many ways a quintessential British Victorian academic, steeped in the classics, and writing books with titles like A Glossary of Greek Fishes (i.e. But he was also a diligent natural scientist, and he became a serious enthusiast of mathematics and physics. And where Aristotle (whom Thompson had translated) used plain language, with perhaps a dash of logic, to try to describe the natural world, Thompson tried to use the language of mathematics and physics.

Russia is testing solar-powered drones that can fly for days at a time above the clouds. If the trial is successful the large glider-like drones could perform some of the same functions as today's space satellites. The model LA-252 Aist will be tested at a height of nine to 13 miles (15 - 21 kilometres) and can be used as a communication device, repeater and Wi-Fi transmitter. Russia is testing solar-powered drones that can fly for days at a time above the clouds. The large glider-like drones could perform some of the same functions as today's space satellites If the trial is successful the large glider-like drones could perform some of the same functions as today's space satellites.

SEE ALSO: Moscow's facial recognition CCTV network is the biggest example of surveillance society yet But when it comes to our personal experience with it, all we usually get is a long, boring, overlooked list of conditions that nobody reads before signing up to Facebook or other social media giants. Do we truly understand what part of our digital footprint is owned by these companies? That's why the Glass Room, which just opened in central London, is important. At first sight, it's just another all-white, sleek, shiny, minimalist pop-up tech store, with massive windows overlooking central London and interactive handsets methodically placed in tactical positions. It bears more than a fleeting resemblance to a famous retail store, which shall remain nameless. Except that once you get inside and start checking out the "products", you're left amused at best, desperately baffled at worst.

Don't throw out your CPUs just yet, but there may be a new way to run your neural networks. In the regular world of computing--whether you're running exotic deep-learning algorithms or just using Excel--calculations are usually performed on a processor while data is passed back and forth to the memory. That works perfectly well, but some researchers have argued that performing calculations in memory itself would save time and energy that is usually used to move data around. And that's exactly the concept that a team from IBM Research in Zurich has now applied to some AI algorithms. The team has used a grid of one million memory devices, pictured above, which are all based on a phase-change material called germanium antimony telluride.