Perspective on the Disconnect in Discourse between Academia and Society on Machine Intelligence

This blog post started as a Facebook post I made recently, but I would like to make this a living document of things I learn about the economic and societal impacts of machine intelligence. Last updated: 10.24.2018


Once in a while it’s nice to peek outside of the academic bubble and see if the newest discoveries have made their way yet to the world at large. Machine intelligence is the single most important scientific endeavor and most powerful technology humans will have ever pursued. As engineers and researchers, it is imperative that we consider and understand the consequences of our designs. Until recently, it wasn’t apparent to me how huge the disconnect is between academia and the everyday person and even the everyday software developer.

Exhibit E2 from McKinsey Global Institute’s 2017 report A Future that Works: Automation, Employment, and Productivity
Full report: https://mck.co/2LHZrf2

It was only a few weeks ago that a recent life event spurred me to read the McKinsey Global Institute’s 2017 report A Future that Works: Automation, Employment, and Productivity. In that report, there’s one chart in particular that stood out to me. Before reading this, I already knew about the threat of automation and had heard rough numbers, but it’s not the same until you see the data for yourself.
The report notes that “while few occupations are fully automatable, 60 percent of all occupations have at least 30 percent technically automatable activities”. However, assuming the 820 jobs that they sampled are representative of the population, then this chart indicates that 25% of jobs are 70% automatable (which I indicated in red)— meaning 1 out of every 4 people are at high risk of losing their job to automation. Perhaps for some jobs, that remaining 30% is too difficult with current technology to completely automate. But given the pace at which the field is moving, that last 30% could be automatable much sooner than most expect.

The report goes on to note that China, India, USA, and Japan “account for just over half of these total wages and employees” with technically automatable activities. If you look at the wealth distributions in these countries, it’s scary to consider what would happen to 1.4B people in India, 1.4B people in China, 127M people in Japan, and 327M in the USA if 25% lost economic opportunities. As of 2018, these four countries represent about 42% of the world’s population. I think these are important factors to consider on all levels — as an individual and as a global community.

Locally in time, the trajectory of society may seem linear. But if you zoom out, the trends indicate that society will require an abrupt paradigm shift and the consequences will be catastrophic. From what I’m seeing, nations aren’t moving fast enough to mitigate those possible futures. Although the report also notes that jobs will be created in the process, not everyone has the opportunity to abruptly acquire a new skill set.

Additional Influential Factors

Additionally, three important factors that the report didn’t mention are latent data reservoirs “dark data”, ease of data collection pipelines, and pace of unsupervised learning. They do mention that new jobs could be created based on previous historical events, “although it was not possible to predict what those new activities and jobs would be while these shifts were occurring”. In previous paradigm shifts, the technology didn’t exist to immediately replace those potential new jobs. Today the data likely already exists and if it doesn’t, then data pipelines can easily be created with a short turn around time for automating those positions. Additionally, with improvements in techniques that require much less labeled data such as pretraining, zero-shot, one-shot, and few-shot learning, and unsupervised learning, those turn around times can be even sooner!

Two more factors later occurred to me, learning on previously unaccessible data types enabled by recent advances in deep learning. In particular, learning on non-Euclidean data such as graph data, which is a subfield referred to as Geometric Deep Learning. Graph data exists pretty ubiquitously from street maps, social networks, protein interactions — essentially anything that can be modeled via relational sets e.g. any data produced from a join operation in MySQL. The other factor that occurred to me is synthetic/simulated data. Improved graphic and simulation algorithms are making it easier to generate high fidelity datasets, which can be combined with transfer learning techniques like Generalizing from Simulation and Learning Dexterity from OpenAI or UC Berkeley’s work on Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World.

When will the effects become more noticeable?

The greatest factor inhibiting the full effects of automation today is adoption. Most companies currently do not yet have the data science or machine learning talent and may not have a full understanding of how to execute large-scale machine learning projects to scale their automation efforts. According to Element AI, there are only “22,000 people worldwide have the skills needed to do serious A.I. research”. However, major universities are doubling down with serious investments in artificial intelligence education. In May 2018, CMU announced their new undergraduate artificial intelligence degree. Five months later, Boston University unveiled their plans (over a decade in the making ) for their new computer science building which will bring a new focus on data science. MIT announced plans for a $1 billion artificial intelligence college. As more institutions for higher education invest in machine intelligence, societal effects will become more globally apparent. More thoughts on this later.

Would really appreciate hearing others’ thoughts on this. Does this chart surprise you? What possible futures do you foresee? How can some of the worst scenarios be avoided?

#12AMThoughts

Further Reading
1. The American Economy is Rigged
2. #Goalkeepers18 | Bill Gates | Is poverty inevitable?
3. World Inequality Report
4. Preparing for the Future of Artificial Intelligence
5. National Artificial Intelligence Research and Development Strategic Plan
6. McKinsey Global Institute, The Age of Analytics Competing in a Data-driven World
7. McKinsey Global Institute, Artificial Intelligence The Next Digital Frontier?
8. McKinsey Global Institute, A Future that Works: Automation, Employment, and Productivity
9. McKinsey Global Institute, Artificial Intelligence Implications for China
10. McKinsey Global Institute, Smartening up with Artificial Intelligence (AI) — What’s in it for Germany and its Industrial Sector?
11. McKinsey Global Institute, Notes from the AI Frontier Insights From Hundreds of Use Cases