Another year has passed since the very first state of AI report, so this year I decided to do (as it turns out not so) quick run down into a couple of topics that I find interesting on the report.
The report is a compilation of 'points of interest' put together by a VC and an Angel investor both focused on AI technologies. The report can be found on slideshare
First things first, they considered the following key dimensions in their report:
- Research: Technology breakthroughs and their capabilities
- Talent: Supply, demand and concentration of talent
- Industry: Large platforms, financing areas of application for AI-driven innovation today and tomorrow
- China: With two distinct Internets, they considered AI in China a different category
- Politics: Public opinion on AI, economic implications and emerging geopolitics of AI
AlphaStar: DeepMind beat a world class Startcraft II player by a whooping 5-0. This represents a huge breakthrough because the AI was running on imperfect information and controlling a huge action space in real time, on top of that, it had to make strategic decisions for long term goals, I personally was watching the match live on Youtube, as a former SC player, it was thrilling.
- How: The network was trained watching humans play (supervised learning) then it created agents and played against itself (Reinforcement Learning).
OpenAi Five: Dota2 bot win rate goes up to 99.4% win rate over 7,000 games with 15,000 live players.
- How: They upped their hardware, this new version now consumes about 800 petaflop/s-days. ### Next
- Creating Creativity: We use rewards in Reinforcement learning to make our algorithm learn, as in real life it's hard to encode rewards, researchers are starting to also give rewards when the algorithm tries new approaches to the problem.
- From Lab to Prod: With the recent release of Facebook's Horizon a Reinforcement Learning platform as an Open Source Software, RL systems optimizations starts to become more widespread. This platform can be used to optimize a set of actions based on the state of the agent at any given moment.
- AlphaFold is at it again: Despite the huge breakthrough last year in protein folding, this year Alpha Fold has significantly outperformed itself.
- Natural Language Processing: This going almost without saying, after the scandal around OpenAI's GPT-2 Model going dark, several labs reported great improvements on this area. Pre-trained models is where it is at guys!
- Common Sense - Some Machines have it, most humans don't: Researchers at NYU trained inferential knowledge to acquire common sense and reason about previously unseen events.
- Growing Interest on Federated Learning: This is like, distributed systems, but for AI. Instead of having a monolithic centralized machine, training is distributed into mobile devices for computation. Google just released their architecture on Federated Learning and how they are using it on GBoard. On top of that, Tensorflow gains a new flavor: Tensorflow Federated. Keep one eye on this one, the health care industry is all over this one.
- Privacy: Companies like Google are starting to give emphasis on data privacy for ML models, Tensorflow Privacy now gives a strong mathematical guarantee that they do not remember your personal details.
- Reading Thoughts: Neural networks can decode your thoughts from brain waves. The technique is still invasive requiring an electrocorticography (for now) but who knows about what can we develop in the next few months???
- Evolutionary AutoML: Evolutionary algorithms are being used for neural network architecture and hyperparameter optimization. The previous technique consisted of RL algorithms.
- The age of Deep Fakes: State of the art GANs are evolving fast, from images to audio, some companies are creating full body pictures with clothes for retail purposes. You guys remember the AI that sounded like Joe Rogan? So, that, and image, and video, and basically everything now.
Top Players (By number of articles on NeurIPS):
- 88% of the researchers at NeurIPS, ICML or ICLR were men.
- 80% of AI professors are men and 75% of students of AI are also men
- Data Labelling jobs are on the rise. ($1.47/h)
- China's studies are growing in average citation rate, despite Europe securing the most number of AI papers published.
- China again: AI courses enrollment is on the rise.
- US, China, UK, Germany and Canada account for 72% of the authors of papers in 2018-2019
- Hiring AI researchers is going down. Perhaps this is a sign to bring this applied research into production environments.
- With big investments, comes ..: Venture Capital in AI grows to reach > $27B/year
- Big Tech investments: Big Tech continues to eat AI-first startups
- Robots taking over: Robotics are on the rise, several processes can be improved like, warehouse management, supply-chain for e-commerce and even robots for building robots.
- Self Driving Road Bump: Self Driving Cars are not the hype at the moment, several players retreated and most companies are missing deadlines.
- Demand forecasting: As more data becomes available in digital form, it becomes increasingly more interesting to use machine learning to forecast demand. With better forecasts, business can prepare supply and reduce waste increasing their profitability. Usage Examples: Travel, Local Businesses, Logistics, Retail and others.
- Reading Machines: With the advances of Natural Language Processing, there's a growing interest in reading and writing machines.
- AI Drugs: The FDA approved 3 AI based diagnostic products and pharma companies are investing in AI-driven drug development.
- Increasing the knowledge gap: AI Patents are on the rise. The most patented field is Computer Vision
- AI Hardware: AI hardware seems to be an interesting field to be in. Several global startups are investing into this technology
- Handheld AI: Samsung, Huawei and Xiomi top the list for performance in AI tasks for mobile devices.
Huawei Hold the Doors: Huawei owns most of the essential patents for 5G tech, likely becoming the key player for building an ecosystem for network providers.
This section is mostly about two surveys, so I'll summarize the results here:
Americans are not in favor of AI for Warfare unless they know someone is doing it
Americans don't know who should decide how AI is developed and deployed
Most people agree that companies should have an AI review board to address ethical decisions
The general public believe that high level AI is just around the corner - 9 years - (considerably sooner than the predictions made by experts)
The majority of Americans trust the US military or academic researchers to develop AI
Facebook was rated the least trustworthy AI developer.
Governance Challenges: Preventing AI-assisted surveillance from violating privacy and civil liberties; Preventing AI from spreading fake or harmful content; Preventing cyber attacks; Protecting data privacy.
Angela Merkel plans to invest 3 billion Euros in AI by 2025
Finland is training 1% of its population in the basis of ML, becoming the first European country to put a national AI strategy in place.
Finland's goal is to become the a world leader in practical AI applications.
Trump signed an executive order creating a program called the "American AI Initiative" with unclear goals.
With Google's exit from project Maven (ML enabled military imagery) a startup was founded and takes over the project and is pushing to employ drones into large-scale conflicts
With recent developments in cameras, mass surveillance is growing in technical sophistication
California approved a bill last year that criminalize the use of bots to interact with a California person "with the intent to mislead".
As stated before, deep fakes are being discussed in the American House of Representatives.
Researchers find racial bias in Amazon's Rekognition
Georgetown Centre for Security and Emerging Technology is the largest US centre focused on AI and policy.
- Alibaba and JD.com have entered the animal and insect husbandry business.
- China R&D spending grows fast but lags in market share
- China is going through increases in industrial automation and job displacement.
- Robots are taking over warehouses in China
- China is publishing more high impact machine learning academic research
- There is a new wave of start-ups for NLP. They will raise over $100M
- Self-driving will stay in R&D.
- Institutions will build AI undergraduate degrees to fill talent void.
- Google will have a major breakthrough in quantum computing hardware, triggering at least 5 startups to do quantum ML.
- Governance and AI will become a bigger topic and one major AI company will make a substantial change to their governance model.
What are your thoughts on their report? What are you most excited about in AI for 2019-2020?