State of #AI 2019 Report

Author: ajit jaokar

I highly recommend  the #StateofAI 2019 report. I have followed this report from By Nathan Benaich and Ian Hogarth

The report is free and you can download it at stateofai 2019

The report is kind of Mary Meeker theme for AI for me i.e. a great reference 🙂 


here are my notes

The full report has lots of slides charts and diagrams. My notes are text only and what was of interest 


AI will be a force multiplier on technological progress because  everything around us today, ranging from culture to consumer products, is a product of intelligence.


The report considers the following key dimensions : Research, Talent, Industry, China(c0onsidered as a distinct category), Politics


Reinforcement learning (RL) 

  • Rewarding ‘curiosity’ enables OpenAI to achieve superhuman performance at Montezuma’s Revenge.
  • StarCraft integrates various hard challenges for ML systems: Operating with imperfect information, controlling a large action space in real time and making strategic decisions over a long time horizon.
  • Quake III Arena Capture the Flag: Human-level performance is achieved by having multiple agents independently learn and act to cooperate and compete with one another.
  • Play-driven learning for robots Training a single robot using play to perform many complex tasks without having to relearn each from scratch.
  • Learning dexterity using simulation and the world real
  • New robotic learning platforms ex UC Berkeley’s Robot Learning Lab created BLUE, a human-scale, 7 degree-of-freedom arm with 7kg payload for learning robotic control tasks.
  • OpenAI used simulation to train a robotic hand to shuffle physical objects with impressive dexterity. The system used computer vision to predict the object pose given three camera images and then used RL to learn the next action based on fingertip positions and the object’s pose.
  • Curiosity-driven exploration How can agents learn to solve tasks when their reward is either sparse or non-existent? Encourage curiosity.
  • Moving research into production environments. Facebook release Horizon, the first open source end-to-end platform that uses applied RL to optimize systems in large-scale production environments, such as Messenger suggestions, video stream quality and notifications. optimisation”).


Natural language processing


A big year in natural language processing

  • Pretrained language models Transfer learning from pretrained language models usher in ‘ImageNet moment’ for NLP.
  • Various research breakthroughs (Google AI’s BERT, Transformer; Allen Institute’s ELMo; OpenAI’s Transformer, Ruder & Howard’s ULMFiT, Microsoft’s MT-DNN)
  • New performance benchmark GLUE
  • A growing interest in federated learning (FL) for real-world products


Machine learning in medicine

  • Expert level diagnosis by AI: Deep learning in medicine: Diagnosing eye disease,  Detecting and classifying cardiac arrhythmia using ECGs 
  • Healthcare: The US FDA cleared 3 AI-based diagnostic products in the last 12 months




  • Robotics in the real world: Cleaning and in-store operations
  • Brain Corp and Walmart are scaling up from an initial 360 robotic floor cleaner trial to add 1,500 more robots. This addresses a $5B commercial floor cleaning equipment market opportunity.
  • Robotics in the real world: From learning to walk to jumping around a parkour course
  • Robots making more robots: Full-stack startups enter the manufacturing market
  • S. factories are installing record numbers of robots 35,880 robots were added to U.S. factories last year, 7% more than in 2017.
  • Amazon is massively scaling its physical fulfilment infrastructure. Amazon rolls out more warehouse robots for fulfilment and sorting. 200k robots (Amazon and third party) in warehouses today, up from 100k announced earlier this year.
  • Amazon has had to make many changes to its warehouses so that its robots know how to navigate around. This includes blocking out sunlight from the ceiling skylights to reduce glare, installing QR codes on the ground and re-orienting air conditioning so that it blows sideways from ground level so as not to push light objects around.


Self driving cars

  • Self-driving cars are now a game for multi-billion dollar balance sheets Cruise sold to General Motors for up to $1B in 2016, Uber spent $457M in 2018, $384M in 2017 and $230M in 2016 on self-driving R&D (including its flying car project).
  • Nonetheless, self-driving mileage accrual in California is still microscopic vs. all drivers Self-driving car companies racked up 0.00066% of the miles driven by humans in California in 2018.


Demand forecasting

  • Demand forecasting Where and how is machine learning being used effectively?
  • Energy: Flood forecasting: Travel: Demand for flights and hotels fluctuates
  • Local businesses: Demand at restaurants, coffee shops or other high street shops is partly dependent on weather and external events.
  • Logistics: Probabilistic models and multi-agent systems can be used to learn how to optimally allocate resources (e.g. fleets of vehicles)
  • Blue Yonder has enabled Morrisons (a supermarket chain in the UK) to fully automate 99% of 20M daily replenishment decisions and in the process improved its profitability and reduced waste.
  • Beyond replenishment, Blue Yonder is applying reinforcement learning to do automated dynamic pricing. This is particularly effective for items that are perishable or subject to consumer trends.



  • Compensation of senior engineers at large tech companies is approaching $1,000,000
  • At the other end of the spectrum, there’s huge growth in $1.47/hour data labelling jobs
  • Plotting countries based on their inflow and outflow of AI talent: Canada, the UK and Switzerland are “platform countries” that both attract foreign talent and export locally-trained talent. The US and Chinese ecosystems are more mature – they see low inflows and outflows.


AI hardware

  • Benchmarking the performance of mobile chipsets for AI tasks Qualcomm’s Snapdragon wins by demonstrating very strong performance and hardware acceleration for both float and quantized neural networks. Benchmarking tasks include classification, face recognition, deblurring, Super-resolution, segmentation and enhancement.


  • Benchmarking the performance of mobile handsets for AI tasks. Samsung, Huawei and Xiaomi top the list whereas Google’s Pixel 3 holds position #22.


  • Pushing compute and competition to the edge Google and NVIDIA throw their hats in the ring to apply AI computation to the 40 trillion gigabytes of data generated from connected devices by 2025. Amazon, on the other hand, launch SageMaker Neo to let developers train ML models on their cloud and export optimised models tailored to specific edge hardware platforms.



  • China is publishing more high impact machine learning academic research
  • Chinese internet companies expand into farming
  • Alibaba and have both entered the animal and insect husbandry business.
  • Chicken farming:In 2016, launched a “running chicken” program to help reduce poverty in Chinese farming regions. Under the program, the company will purchase any free-range chicken that runs over one million steps for three times the going market rate. Now, has expanded the program to integrate AI tools across the husbandry workflow.
  • Pig farming: In a collaboration between Dekon Group, Tequ Group and Alibaba Cloud, a computer visionand voice recognition system is used to identify individual pigs via numbers tattooed on their flanks and to monitor vulnerable piglets for squeals of distress. By 2020, Dekon wants to breed 10M pigs per year.
  • Robots are driving automated warehousing in China
  • Chinese groups own the most patents, but only 23% were “invention patents” in 2017
  • Chinese inventors let the majority of their patents lapse 5 years after they’re granted


Policy and public attitude

  • Public Attitudes to AI: Warfare and double standards. Overall, Americans are not in favour of developing AI technology for warfare, but this changes as soon as adversaries start to develop them.


  • Public Attitudes to AI: Governance. Who should decide how AI Is developed and deployed? The majority of Americans don’t know.


  • Should companies have an AI review board that regularly addresses corporate ethical decisions?


  • Public Attitudes to AI: high level machine intelligence is just 9 years off…High-level machine intelligence was defined as machines able to perform almost all tasks that are economically relevant today better than the median human (today) at each task.


  • Top perceived AI governance challenges: Preventing AI-assisted surveillance from violating privacy and civil liberties; Preventing AI from being used to spread fake and harmful content online; Preventing AI cyber attacks against governments, companies, organizations, and individuals.


  • Protecting data privacy.


  • AI Nationalism: ex “AI made in Germany” Plan announced to invest 3 billion Euros by 2025. – other countries following same


  • Rather than competing with China and the US, Finland aims to occupy a niche as world leader in practical applications of AI, says Economy Minister Mika Lintilä.


  • Europe aims to differentiate by focusing on “ethical AI” and its reputation for “safe and high-quality products”.


  • Trump’s AI plan and export controls.


  • Google canceled its Maven contract with the US Defense Department after 4,000 employees protested.


  • With research breakthroughs in NLP come dual use concerns.


  • As machines get better at reading and writing there is increasing scope for fraud (scalable ‘spearfishing’ attacks over email for example) and computational propaganda.


  • Concerns over this have caused OpenAI to run an experiment in “responsible disclosure” and only share a smaller version of their latest language model, GPT-2 to avoid misuse.


  • New challenges: DeepFakes hit the political agenda


AI Industry trends



  • AutoML
  • State of the art in GANs continues to evolve: From grainy to GANgsta
  • Analysis of 16,625 AI papers over 25 years shows immense growth in publication output with machine learning and reinforcement learning being the most popular topics
  • 10x more papers annually over the last 10 years. Over 50% of papers are about machine learning.
  • Google continues its dominance at NeurIPS 2018, a premier academic AI conference
  • Google tops list of the most productive organisations measured by research paper output.
  • Similar to the complex electronics supply chain (for example Foxconn), there has been massive growth in ‘data labelling factories’ for AI applications.
  • New $1 billion investment in computing & AI at MIT: Shifting gears for new generations. Anchored by a $350M gift by Blackstone CEO Stephen Schwarzman, the new College of Computing will reorient MIT towards injecting AI education to all fields of study. It provides 50 new faculty positions, doubling MIT’s academic capability in the field.
  • Gender diversity of AI professors and students is still far off being on an equal footing A review of papers published at 21 machine learning conferences by 22,400 unique authors: Only 19% of academic authors and 16% of industry authors were women.
  • 44% of authors earned a PhD from the US, 11% from China and 6% from the UK.
  • Five countries — the US, China, the UK, Germany and Canada — accounted for the employment of 72% of the authors
  • Global venture capital investments in AI themes grow at a clip to reach >$27B/year Almost 80% more capital invested in FY18 vs FY17 with North America leading the way at 55% market share.
  • Robotic process automation: An overnight enterprise success (15 years in the making)
  • Reading machines are improving and proliferating The breakthroughs in natural language processing highlighted in the Research section of this report are starting to be applied to industries where there are either large amounts of text to be processed or where there is substantial financial return from processing text faster.
  • Computer vision is the most popular area for patents. Within computer vision the most popular area is biometrics (applications related to biological data).
  • Big tech companies monetise cloud computation, but not their hosted AI services
  • 5G as the backbone of ubiquitous connectivity and AI computation. 5G offers the potential for much faster and more stable information transmission. The organisation or country that owns 5G will set the standards for the rest of the world.



Go to Source