It is well understood that the performance of machine learning methods is heavily dependent on the choice of data representation (or features) on which they are applied. The rapidly developing field of representation learning is concerned with questions surrounding how we can best learn meaningful and useful representations of data. We take a broad view of the field, and include in it topics such as deep learning and feature learning, metric learning, kernel learning, compositional models, non-linear structured prediction, and issues regarding non-convex optimization. Despite the importance of representation learning to machine learning and to application areas such as vision, speech, audio and NLP, there was no venue for researchers who share a common interest in this topic. The goal of ICLR has been to help fill this void.

Let us all raise a glass to AlphaGo and mark another big moment in the advance of artificial intelligence and then perhaps start to worry. AlphaGo, Google DeepMind's game of Go-playing AI just bested the best Go-playing human currently alive, the renowned Lee Sedol. This was not supposed to happen. An artificial intelligence capable of beating the best humans at the game was predicted to be 10 years away. But as we drink to its early arrival, we should also begin trying to understand what the surprise means for the future – with regard, chiefly, to the ethics and governance implications that stretch far beyond a game.

We return to the question of terminology that we started this post with. Our feeling is that the term "artificial intelligence" has been used in so many ways that it is now confusing. People use AI to refer to all three approaches described above, plus others, and therefore has become almost meaningless. The term "machine learning" is a more narrowly defined term for machines that learn from data, including simple neural models such as ANNs and Deep Learning. We use the term "machine intelligence" to refer to machines that learn but are aligned with the Biological Neural Network approach. Although there still is much work ahead of us, we believe the Biological Neural Network approach is the fastest and most direct path to truly intelligent machines. This blog entry was modified on Thu Mar 24 2016 to clarify the timing of neural network research.

During my last interview I had a great talk with Daniel McDuff. Daniel's research is at the intersection of psychology and computer science. He is interested in designing hardware and algorithms for sensing human behavior at scale, and in building technologies that make life better. Applications of behavior sensing that he is most excited about are in: understanding mental health, improving online learning and designing new connected devices (IoT). Listen to more about why it is important to collect data from much larger scales and help computers read our emotional state. Key Learning Points: 1. Understanding the impact, intersection, and meaning of Psychology and Computer Science 2. Facial Expression Recognition 3. How to define Artificial Intelligence, Deep Learning, and Machine Learning 4. Applications of behavior sensing with Online Learning, Health, and Connected Devices 5. Visual Wearable sensors and heart health 6. The impact of education and learning 7. How to build computers to measure phycology, our reactions, emotions, etc 8. Daniel is building and utilizing scalable computer vision and machine learning tools to enable the automated recognition and analysis of emotions and physiology. He is currently Director of Research at Affectiva, a post-doctoral research affiliate at the MIT Media Lab and a visiting scientist at Brigham and Womens Hospital. At Affectiva Daniel is building state-of-the-art facial expression recognition software and leading analysis of the world's largest database of human emotion responses. Daniel completed his PhD in the Affective Computing Group at the MIT Media Lab in 2014 and has a B.A. and Masters from Cambridge University. His work has received nominations and awards from Popular Science magazine as one of the top inventions in 2011, South-by-South-West Interactive (SXSWi), The Webby Awards, ESOMAR, the Center for Integrated Medicine and Innovative Technology (CIMIT) and several IEEE conferences. His work has been reported in many publications including The Times, the New York Times, The Wall Street Journal, BBC News, New Scientist and Forbes magazine. Daniel has been named a 2015 WIRED Innovation Fellow.

At the GPU Technology Conference, Nvidia unveiled the Tesla P100, the latest addition to Nvidia's Tesla Accelerated Computing Platform (TACP). The accelerator unit is being marketed as the most advanced hyperscale datacenter accelerator ever built – with a claimed 12x improvement over the previous Maxwell architecture, thanks to the new Pascal architecture. Designed to provide the equivalent performance of hundreds of general purpose CPUs in a much smaller package, and with significantly lower opex costs, Nvidia is targeting the next-gen datacenter use cases, which consist largely of artificial intelligence applications – which require very different compute resources than most current datacenters can provide. Cloud computing and the supercomputing that powers dense data analytics are very important for the progression of the Internet of Things (IoT). With the image-recognition that will power computer visions, smart grid management, smart city operations, and the massive amounts of sensor data that need to be crunched to realize more efficient business practices, systems like Nvidia's provide a very capable alternative to gigantic arrays of general purpose compute resources in datacenters.

A discussion of three major approaches to building smart machines - Classic AI, Simple Neural Networks, and Biological Neural Networks - and examples as to how each approach might address the same problem.

Elon Musk has not been shy about his concerns over artificial intelligence turning evil. So it wasn't a surprise in December when Musk announced the formation of OpenAI, an open-source, non-profit focused on advancing "digital intelligence in the way that is most likely to benefit humanity as a whole." That's all well and good, but not much has been revealed about what exactly OpenAI is working on. OpenAI's co-founder and CTO told Tech Insider that OpenAI is primarily focusing on advancing machine learning, which is the technology that enables computers to learn how to complete tasks through experience. Specifically, the company is focusing on two key types of machine learning that every major tech company is investing in right now.

Seismic launched their own collaboration platform called Workspace and claim to be "the only sales enablement platform with every essential capability in place now." The new cloud forum enables enterprise users to collectively gather, share and access content and add comments, annotations, feedback and commentary to their sales content. Amazon acquisition of deep learning startup Orbeus Inc, underlining their efforts to apply AI based deep learning techniques in their delivery systems and cloud computing business. Orbeus has developed a photo recognition technology based on neural networks named ReKognition. Conversocial, a social customer care platform launched its Channel API and announced initial integration partnerships with social Intelligence, monitoring and analytics companies Synthesio and Brandwatch.

For a while (at least several months since many people began to implement it with Python and/or Theano, PyLearn2 or something like that), nearly I've given up practicing Deep Learning with R and I've felt I was left alone much further away from advanced technology… But now we have a great masterpiece: {h2o}, an implementation of H2O framework in R. I believe {h2o} is the easiest way of applying Deep Learning technique to our own datasets because we don't have to even write any code scripts but only to specify some of its parameters. That is, using {h2o} we are free from complicated codes; we can only focus on its underlying essences and theories. With using {h2o} on R, in principle we can implement "Deep Belief Net", that is the original version of Deep Learning*1. I know it's already not the state-of-the-art style of Deep Learning, but it must be helpful for understanding how Deep Learning works on actual datasets. Please remember a previous post of this blog that argues about how decision boundaries tell us how each classifier works in terms of overfitting or generalization, if you already read this blog.

Jon Barker: Jon Barker is a Solution Architect with NVIDIA, helping customers and partners develop applications of GPU-accelerated machine learning and data analytics to solve defense and national security problems. He is particularly focused on applications of the rapidly developing field of deep learning. Prior to joining NVIDIA, Jon spent almost a decade as a government research scientist within the U.K. Ministry of Defence and the U.S. Department of Defense R&D communities. While in government service, he led R&D projects in sensor data fusion, big data analytics, and machine learning for multi-modal sensor data to support military situational awareness and aid decision making. He has a Ph.D. and B.Sc. in Pure Mathematics from the University of Southampton, U.K.