If you haven't heard about the artificial intelligence (AI) machine-learning (ML) craze that uses deep neural networks (DNN) and deep learning (DL) to tackle everything from voice recognition to making self-driving cars a reality, then you probably haven't heard about Google's new Tensor Processing Unit (TPU), Intel's Lake Crest, or Knupath's Hermosa. These are just a few of the vendors looking to deliver platforms targeting neural networks. The TPU contains a large 8-bit matrix multiply unit (Figure 1). It essentially optimizes the number-crunching required by DNN; large floating-point number-crunchers need not apply. The TPU is actually a coprocessor managed by a conventional host CPU via the TPU's PCI Express interface.

It's just over a year since Google's DeepMind unit stunned the world when its AlphaGo AI beat Go legend Lee Se-dol 4-1 in a five-game match; the result demonstrated mastery of a feat that had eluded computer scientists for decades and sparked a flood of new interest in the field of artificial intelligence. But there was one possible "gotcha" that Go devotees could hold onto: Lee Se-dol was once, but is no longer, quite considered the greatest player on the planet. That distinction is now considered to belong to Ke Jie, a 19-year-old Chinese player ranked number 1 worldwide. A professional since the age of 10, Ke has beaten Lee several times in high-profile matches in recent years, including three finals victories in the three months leading up to Lee's AlphaGo match. And next month, Ke will get his own showdown with DeepMind's AI.

I love motivational examples in teaching complex ideas! I use this simple little video to teach Deep Learning to my students. When we consider Deep Learning, we think of ideas like teaching a computer to recognize images of cats using Deep Learning OR training a computer to play pacman using Deep Learning. You let the Deep Learning system iterate with many examples and in each case, you tell the computer using a classifier if its interpretation was correct or not (aka is it a cat or not, Pacman scores, etc.) Now watch the video - video link is lipstick robot. I see the click at the end of the step as a classifier.

Deep learning is currently a central pillar of the Twitter experience. For most of the social network's history, users saw tweets in their timelines from people they follow in the chronological order in which they were posted. But about a year ago, Twitter changed to a ranked timeline. Similar to how Facebook operates its ranked news feed, this new approach assesses people's tweets for content and surfaces them in a user's timeline based on how strongly the algorithm thinks they will appeal to the user. A lot of deep learning is involved in this process, from natural language processing to image recognition and description.

Figure 2: Our model was trained using a combination of exogenous variables, including weather (e.g., precipitation, wind speed, and temperature forecasts) and city-level information (e.g., trips in progress at any given time within a specific geographic area, registered Uber users, and local holidays or events).

The theatrical release of James Cameron's sci-fi film Terminator 2, featuring Arnold Schwarzenegger as a cyborg with a computer brain, had a crucial scene deleted. The scene, part of the extended release of the movie, shows young John Connor and his mother opening up the head of the cyborg to switch its computer brain from "read only to "learning" mode. The cyborg (Schwarzenegger) then picks up human values and mannerisms as the movie progresses. For movie buffs, the deleted scene is worth seeing for special effects and also to catch a glimpse of Linda Hamilton (playing John's mother Sarah Connor) with her twin sister Leslie playing her image in a mirror. In the theatrical release, where the scene is omitted, the cyborg just tells John that its brain is a "neural-net processor, a learning computer", without mentioning any on/off options. That was back in 1991. Today, in 2017, a learning computer is much more of a reality. While artificial intelligence (AI) and machine learning (ML) concepts have been around since the 1940s and 1950s (See ABC of AI, ML and Deep Learning), the availability of huge amounts of data is making the difference now. A "learning computer" does not need to travel back in time -- like in the movie -- and many are solving real problems in India. For example, in healthcare, ML is helping oncologists sift through huge amounts of cancer cases and suggesting preferred treatment; in education it is predicting who might drop out of school; and in fashion it is forecasting colours that can dominate the next season. Retail, transportation and financial services have adopted ML in different forms. The "learning switch" is turned on in India. "Every large organisation was sitting on data.

Many vendors like Cisco and Intel are proponents of Edge Processing (also called Edge computing). The main idea behind Edge Computing is to push processing away from the core and towards the Edge of the network. For IoT, that means pushing processing towards the sensors or a gateway. This enables data to be initially processed at the Edge device possibly enabling smaller datasets sent to the core. Devices at the Edge may not be continuously connected to the network.

For example, some Indian researchers used CRF to extract key words from medical texts and they had good features and large enough training sample, but they obtained quality not more than 0.4 (F1-measure). On real data they would hardly obtain such quality, while Stanford NER shows quality not more than 0.81 (F-measure) given it has perfectly selected training features and it was trained on larger corpora (CoNLL, MUC-6, MUC-7 and ACE) Some Spanish and Russian researchers compared HMM and CRF in NER task for medical texts on JNLPBA corpus (18546 sentences with 109588 named entities). They obtained interesting results: HMM had higher recall ( 4-7% depending on the type of entity) while CRF had higher precision ( 4-13% depending on the type of entity). According to one master thesis, linear-chain CRF operated very well on extracting time expressions from Russian text.

As a Windows user interested in learning to use GPUs for deep learning I have been frustrated with the available software and books for a couple of years. Most books seem to have been written on Linux systems and then you are told that the code should work fine with Windows - which is not always so. Also several deep learning systems have not been set up for Windows yet. So when I first read about TensorFlow I decided to give it a try: I got this book and one other. Final evaluation first: I gave it 4 stars because the code worked for me (more accurately: much, not all of the code worked), but the situation was better than other books hence 4 stars.

Multiagent environments where agents compete for resources are stepping stones on the path to AGI. Multiagent environments have two useful properties: first, there is a natural curriculum -- the difficulty of the environment is determined by the skill of your competitors (and if you're competing against clones of yourself, the environment exactly matches your skill level). Second, a multiagent environment has no stable equilibrium: no matter how smart an agent is, there's always pressure to get smarter. These environments have a very different feel from traditional environments, and it'll take a lot more research before we become good at them. We've developed a new algorithm, MADDPG, for centralized learning and decentralized execution in multiagent environments, allowing agents to learn to collaborate and compete with each other.