Inside | Real news, curated by real humans
Inside AI

Inside AI (Dec 22nd, 2019)

Happy Sunday and welcome to InsideAI.  I'm Rob May, a General Partner at PJC focused on early stage AI companies.  This week, I have an interview with my friend Yaneer Bar-Yam, a complex systems mathematician who runs the New England Complex Systems Institute.  But before we get into that, here are the most popular articles from the weekday version of InsideAI:

Developers created an AI that automatically produces "hedcuts" – the style of portrait used by The Wall Street Journal for its columnists. The Journal began using hedcuts - which are drawings based on the so-called "stippling" and "hatching" methods - back in 1979. While a single portrait can take about three to five hours to create, the publication's new AI can apparently develop one from scratch. Boing Boing points out that the machine ran into some obstacles along the way, including not understanding baldness and overfitting, which caused it to produce some "terrifying monstrosities" (like the hedcuts seen below). - WSJ

Image editor Pixelmator unveiled a machine-learning feature that can sharpen and enhance low-res images. The company, which competes with Photoshop, says its “ML Super Resolution” can scale an image up to three times the original resolution without adding pixelation or blurriness. Pixelmator’s creators say the algorithm is 5MB in size, allowing it to run on users’ devices, and was trained on 15,000 samples of low-resolution and high-resolution images, where it learned to "fill in the gaps" between images pixel by pixel. After tests, The Verge reports that it's the "best commercial super-resolution tool" it's seen so far, although it does have caveats. - THE VERGE

Health care startup Paige - which is using deep learning networks to help detect cancer - has raised $45 million in a new funding round led by Healthcare Venture Partners. Breyer Capital, Kenan Turnacioglu, and other funds also participated. New York-based Paige, formerly Paige.AI, emerged from stealth last year from its beginnings at the Memorial Sloan Kettering Cancer Center, where it secured access to 25 million pathology slides and intellectual property related to AI computational pathology. The company uses computer vision trained with clinical imaging data to map the pathology of cancer and better diagnose the illness. Paige says it will use the capital to seek FDA clearance for its products in areas like prognostic capabilities and biomarkers. - TECHCRUNCH

Joelle Pineau, a machine-learning scientist at McGill University and Facebook, wants more AI researchers to open up their code and methods because she says it could lead to better reproducibility. Pineau launched a reproducibility competition - challenging researchers to re-create each other's work - at this month's Conference on Neural Information Processing System in Canada. Organizers encouraged attendees to submit their code and review a special checklist for submitted papers, which asks for details of models and other characteristics. The efforts could be paying off: A year ago, half of accepted NeurIPS papers contained a link to code, compared to 75 percent this year, Pineau noted. - NATURE.COM

I became fascinated by complex systems many years ago when I read the book Logic of Failure.  So two years ago when I found the complex systems course offered by the New England Complex Systems Institute, I signed up.  That's where I met Yaneer Bar-Yam, one of the top experts in complex systems mathematics.  Over the past year, Yaneer and I have discussed AI and complex systems, and the fact that the latter can solve some types of problems that AI can't.  I've seen it proven out in some of the results of the consulting work NECSI does for industry - to solve the hardest problems that companies can't seem to solve using normal statistical tools. 

As we look to 2020, I've been thinking about the limitations of AI, where it may go next, and what might be beyond AI, so below is an interview with Yaneer that can give you exposure to some ideas in that direction.  If you find this interesting, NECSI has a great introduction to complex systems on their website.

1.  What is your background, and wow did you first get into complex systems?

I am a physicist and my early work was studying material properties. I began to work on complex systems because there seemed an opportunity to think about topics ranging from biological molecules to human cognition to social systems. Instead of thinking about how materials are made of atoms, we could think about how biological, psychological or social systems were made of their parts, whether atoms, neurons, or people.

2.  When and why did you start NECSI?

There was a gap between what faculty at universities were considering and what I and a few others were doing. First, because the underlying ideas were different, second because the boundaries between academic subjects/departments didn’t work for this area. Students were working with me from several different departments and they didn’t have a proper home.

3.  One of the things I found fascinating when I took the NECSI course is that, for many problems where we are looking for a simple answer, the actual complex systems answer is "it depends".  What types of things does "it depend" on, and how do you get people to grasp that there aren't always simple answers?

When people are trying to understand or solve a problem they make simplifying assumptions and don’t check if they are correct. There are many different ways that this can fail. One of the most important ways we simplify things is to assume that the answer doesn’t depend on the context. What is right for people in one place in the world is right for people in other places, or with different backgrounds: The right product, the right treatment, the right way to speak to them, and so on. When we recognize this is a problem, we may get overwhelmed with the many different ways things might be different. If there isn’t a solution to being overwhelmed we might as well go back to the simplifying assumption. The trick is to focus on the specific ways that do affect the conclusion and ignore those that don’t matter. That is the path forward and how people can grasp that we can go beyond an overly simple approach.

4. AI and Machine Learning have been on the rise for a few years, and can solve many problems other approaches can't.  What types of problems can complex systems mathematics solve that might be beyond the current capabilities of Machine Learning?

Machine Learning maps data onto a set of actions, which may be informing people or may be direct action, i.e., automation. In order for this to be done, it requires that we have enough data to map out all of the cases that map onto those actions. What happens if we don’t have enough data? Complex systems mathematics enables us to develop a “deep structure” of the system which is a counter point to the “deep learning” of machine learning. That deep structure enables us to identify the right actions even when there isn’t enough data for machine learning.

5.  Have AI and ML made complex systems approaches better?  Are there cases they can work together?

We consider AI and ML as among the tools of complex systems science. We were using neural networks and machine learning before it became a thing in the past few years. The advances that gave rise to the excitement about new applications are powerful and useful, and we make use of them when the are the right way to address a particular problem. These are some of the building blocks for how we develop solutions based on complex systems science.

6.  Causality has become a hot topic in AI.  Is this an area where your work can help?

Yes. Once we have the deep structure of the problem, we can build causal models and validate them. This is where conventional statistics relies upon correlations to characterize dependencies in a system, but complex systems science can build models that are inherently causal. Causal models then enable predictions and “what if” scenarios that are much more powerful than data driven analytics by itself.

7.  NECSI is a non-profit, but you do a fair amount of work for industry.  What types of customers come to NECSI?

In recent years we have had the opportunity to partner more with corporations. Examples include, technology companies developing platforms for communication and coordination, and industrial companies on supply chain optimization and customer experience. We have also studied the dynamics of global trade patterns to evaluate differential opportunities. Our customers have in common a need for solving strategic problems that will have a high impact on their organization, either in responding to operational challenges or to growth opportunities. Also they are willing to look beyond approaches they have used historically. 

8.  Over the next decade, do you see AI and complex systems merging more closely together, or drifting further apart? 

AI today is limited to be mostly a tactical tool for cases where the data is large enough for answering specific types of questions. Over time, I expect that we may incorporate into AI more of the tools available in Complex Systems Science, so that the concept of intelligence continues to grow. Alternatively, we will simply recognize the existence of a variety of analytic tools and consider AI to be one of them.

Thanks for reading, and see you next Sunday with my 2020 areas of focus for AI!


Subscribe to Inside AI