Recently 2017 came to a glittering end and as we look back through the lens of technology, the winner was probably Artificial Intelligence aka AI. It received tremendous success as much as some of the industry leaders commented that 2017 was the ‘Year of AI’. This write-up is an attempt to collate the achievements under the academic and industry.
Starting off with academics, the sheer volume of papers published is increasing every year. To give you some statistics, in 2017 it was 9 times more than 1996. Similarly, the number of enrollments for AI courses went up by almost 11 times in 2017 compared to early 2000.
In terms of research, Stanford University developed an AI system that can detect life-threatening irregular heartbeats by quickly sifting through hours of heart rhythm data and with cardiologist level accuracy. They have also developed a system that can predict court decisions better than legal scholars, even with less amount of information at its disposal. The Florida State University developed an AI algorithm that can predict whether someone will attempt suicide as far off as two years into the future with up to 80 per cent accuracy. Researchers from Middlesex developed a system called VALCRI (Visual Analytics for Sense-Making in Criminal Intelligence) which can help solve crimes by taking over the laborious task of analysing clues and finding links that human investigators might have missed. The researchers at MIT have devised an AI technique which will go through research papers and deduce recipes for making custom/particular material. There are many more examples which can be cited here. There was also tremendous growth in the number of MOOC courses and specialization launched for learning. The foremost of which was deeplearning.ai launched by Andrew Ng. There are several other launched by edX, Udacity, Udemy too. Overall, we may state that the academic research work in the field of AI gained a lot of prominence and the focus on training to prepare the workforce for AI also gained popularity in 2017.
Now moving on to the industry, we would be analysing under a couple of heads Applications, Software, Investment and Hardware. Let’s start off with the building blocks: Hardware Systems.
There were competition & collaboration between companies to produce chips and processors geared towards faster processing, more scalable models, optimizes memory and provides more computational power. The AI chips transcended the computers and landed on top of mobile and other devices boards too. Intel launched the Nervana™ Neural Network Processor which it claims to be the first Neural Network silicon designed for enterprise deployment. It also launched the Movidius™ Neural Compute Stick, which it claims to be the world’s first USB-based Deep Learning inference kit and AI accelerator. Nvidia launched the most powerful GPU named Titan V, which is based on its advanced AI chip architecture named Nvidia Volta. Tesla also announced that it would be launching AI Hardware soon. Coming to Google, it already had TPU under its hood from 2016 & it introduced the 2nd generation of TPU and launched the Cloud version: Cloud TPU in 2017. It also launched the 1,000 Cloud TPUs available at no cost to ML researchers. IBM too announced the AI chip named Power9, which it claims will have 4 times better performance for frameworks like Caffe. Moving on to the smaller devices, one of the most significant achievement was on device AI acceleration. Apple in its new iPhone series had embedded a new chip named A11 Bionic which included GPU units and a dedicated Neural Network hardware called ‘Neural Engine’ which can perform 600 Billion operations in 1 sec. Qualcomm announced its next generation Snapdragon 845 mobile platform series which is planning to move the AI from cloud to the chips (device). This will have the Qualcomm® Hexagon™ 685 DSP which supports the on-device AI processing. In terms of other devices, two significant mentions would be the announcement from Microsoft for use of AI Chip in Hololens and other devices going forward. The last and not the least would be Amazon’s announcement of DeepLens which it claims is ‘the World’s First Deep-Learning Enabled Video Camera for Developers’. There were significant developments by the start-ups in this area too but the challenges are that the big bothers are already in the market and it may take a lot of time before they can get the chip to the market as noted by this writeup from MIT.
First let’s start with the AI programs which can code itself. The important mentions are AutoML from Google, where in a blog in May,2017 they announced that ‘a controller neural net can propose a “child” model architecture, which can then be trained and evaluated for quality on a particular task’. It further followed it up with the announcement of an architecture named ‘NASnet’ which achieved about 82.7 % accuracy on ImageNet classification. Microsoft also came up with DeepCoder in collaboration with Cambridge, which gained considerable press. There were a lot of frameworks/launched which was launched for AI this year notable among them are CoreML from Apple, which is a mobile machine library, Uber released Pyro, a probabilistic modelling language, which according to Uber unifies ‘the best of modern deep learning and Bayesian modeling.’ Amazon and MS jointly launched Gluon an open source library in Apache MXNet which ‘is an effort to improve speed, flexibility, and accessibility of deep learning technology’. Deepmind also open sourced its framework for building complex neural networks called Sonnet on top of Tensorflow. A new open standard for interoperability for deep learning models was launched by Amazon, Facebook and MS called Onnx (Open Neural Network Exchange). This enables interoperability between Apache MXNet, PyTorch, Caffe2, Tensorflow, and Microsoft Cognitive Toolkit. Other mentions are OpenAI’s Roboschool for robot simulation and Baselines for implementations of reinforcement learning algorithms, Tensorflow Agents promises efficient Batched Reinforcement Learning by extending the OpenAI gym interface and Facebook’s launched ELF platform for real-time strategy (RTS) games. There were some deep learning web frameworks which was launched 2017, notable are deeplearn.js by Google, MIL WebDNN which runs trained DNN Model on Web Browser and TensorFire which promises ‘in-browser flaming-fast gpu-accelerated deep-learning’.
The areas of application were diverse and the results exciting, with concepts ranging from extra-terrestrials to core functions of the human body. AI helped us to identify a distant solar planet about 2500 light years away. This was announced by NASA recently in partnership with Google. We discussed about a couple of Medicine application right in the beginning of this write-up, all of them deal with nuances of the human body. There were a couple of key applications/models in the field of Generative Adversarial Networks which were promising. There was obviously the much-celebrated application of AI, when AlphaGo defeated the world champion of Go. Libratus AI developed by Carnegie Mellon defeated professional poker players too. There was Sophia, a humanoid robot designed by a company in Hong Kong, that was granted citizenship in Saudi Arabia, which caught the attention of the news media for a couple of days. There was also this fantastic piece of work ‘Using Deep Learning and Google Street View to Estimate the Demographic Makeup of the US’ where using car images captured by Google Street View cameras, the demographic details of US cities were calculated.
Overall the tech giants spent an estimated $40-50 billion on AI through 2016-17. There were a lot of acquisitions, some of the key ones are Microsoft’s acquisition of deep learning start up Maluuba, Google acquired Kaggle, AIMater & HalliLabs, Facebook bought AI assistant startup Ozlo, Samsung acquired Fluently, Apple acquired Realface, Lattice, AWS got cyber-security start-up Harvest.ai and Blink, Spotify acquired Niland which optimizes music search, Ford acquired Argo AI, Intel acquired Mobileye, Baidu acquired Kitt.ai etc. In the start-ups market, AI landed with about $12 billion of equity funding, 350+ equity deals and about 100+ start-ups came into prominence globally.
To conclude, 2017 was a phenomenal year for AI in terms of advancement in terms of research, hardware, software, investments and application. There are more and more people joining this field from various domains and projects are getting open sourced which is a welcome change as it fosters collaboration. There were a couple of hick-ups along the way as well, but they are expected in a field where there is a lot of activity and upheavals. 2018 promises to be a year with wider adoption of AI in our day to day lives. Virtual assistants would be more widely used, the voice enabled devices may start talking to each other, there would be higher usage of biometrics for security/payments, the medical community may start using this as a tool for diagnosis and there would be novel ways of using AI as a tool for business, governance and information technology industry.