Recently, I saw a demo of some innovative machine learning algorithms used to fly autonomous drones. As I discussed the very cool details, I reflected on how rare it is that I hear or read about similarly cool artificial intelligence/machine learning (AI/ML) that materially transformed (for the better) a financial services firm. This is the space where I, as a machine intelligence (to be defined later) researcher, spend most of my time (when I am not buried in management demands.) Despite nearly a decade of promises from the AI/ML communities, financial services, generally and insurance, specifically are yet to be transformed by these promising technologies. In the last year, this failure has been reflected in several studies (see gartner.com for a sample.) Estimates of AI/ML project failures range from 80% and up (my personal view as someone who bridges the AI/ML & financial services communities is something like 95%). I have attached at the end of this blog an image from gartner.com showing how many of the relevant AI/ML technologies are now in the "trough of disillusionment." The picture looks bleak.
This blog is the first of many I plan to write related to discussions we are having within the Swiss Re Institute (SRI), among SRI's academic & technology partners, and Swiss Re's clients. SRI wants to broaden this discussion of how to deploy successful, transformative enterprise machine intelligence across financial services. As an aside, SRI is sponsoring an AI & insurance track at the Applied Machine Learning Days to be held from January 25 to 28 at EPFL in Lausanne, Switzerland. This track will be another forum to deepen this debate.
As those of us with advanced academic training know, addressing research problems usually starts with definitions & taxonomies in order to define clearly a target research space with relevant hypotheses. I started this blog referring to AI/ML, which are now common acronyms to describe how machine algorithms infiltrate our lives. For general/overarching conversations, I prefer the term machine intelligence (MI) (which encompasses conventional curve-fitting, various types of machine learning, and artificial intelligence—broadly defined) and am fighting a losing battle to reserve AI & ML for more specific discussions. This said, definitional distinction matters for figuring out how to make MI relevant for financial institutions (FIs). If we can clarify proposals with more specific language, we will collectively do a better job of matching particular solutions to specific use cases. Since I work within the insurance industry, my examples will skew toward insurance; however, this debate will benefit from cross-fertilization arising from different kinds of FIs (e.g., banks, asset managers, re/insurers, etc.) speaking with each other and with their technology partners.
At this point, you may ask "why should I care about language & definitions?" The answer relates to better articulating two (of several) recommendations for how to increase substantively the probability of deploying successful, transformative, end-to-end enterprise machine intelligence:
1) Clarifying business use cases & relevant project plans for cross-functional MI implementation teams.
2) Refining roles, responsibilities, background, & expectations for cross-functional MI implementation teams.
Let's start with the term intelligence: Here we see an example of why English benefits from being such a rich mix of Latin, Old/Middle English, and Old French.
The core meaning starts with the Latin word intelligere, which means "to understand." This spilled over to 12th century Old French, which defined the word as "faculty of understanding." And then by the late 14th century, middle English expanded this word's applicability in a philosophical direction-- "… capacity for comprehending general truths."
What's interesting is that by modern times the definition has morphed from understanding/comprehension to "the ability to acquire and apply knowledge and skills." (lexico.com) This modern definition of intelligence implies an adaptivity and flexibility not in the original notion of to understand. Here we see an important bifurcation in the machine intelligence space:
· "Dumb" algorithms that just regurgitate "understood" patterns in a large set of data e.g., deep learning applied to facial recognition. These types of algorithms reflect more of the original definition of intelligence hundreds of years ago.
· "Smarter" algorithms that develop some ability to acquire and then apply—not necessarily always in line with what has been seen before—knowledge & skills. Here the modern definition of intelligence encompasses approaches likely to be more successful in FIs.
A deep & tortuous philosophical continues to rage with respect to what it means to know (theory of epistemology). I don't want to fall down that hole—rather, I want to emphasize that much of the success in the MI space is primitive "understanding" that requires loads of data & little change in the underlying data generating process. Few use cases in financial services fit into this category.
All the above said, "simple" machine intelligence such as regression, non-parametric statistics, etc. have seen success in enterprise financial services deployment for risk assessment, certain types of risk management, pricing, and short-term forecasting. But I don't think this type of machine intelligence (which we used to call "statistics") is what C-suite executives think they are getting when they pay for "AI/ML" projects. Intelligence as currently defined reflects a characteristics where humans still have an advantage over machines. This may change—not very quickly. As I discuss the notion of learning below, I highlight a few areas of research where we may see substantive progress in deploying "smarter" algorithms.
I will forgo a linguistic detour with respect to the word "machine" and mention just one salient point in the enterprise MI context—machine intelligence to be successful in financial services needs the "machine" to be defined in terms of an end-to-end process integrated into regular business workflow. This means the "machine" is not just the algorithm. Many (if not most) failed enterprise MI deployments are declared successful in the narrowly defined algorithmic space. An algorithm, by itself, is not a useful machine for an enterprise.
"Learning" is another word where etymology does not add insight. The issue here has to do with whether the "learning" constitutes "dumb" pattern regurgitation or whether some degree of creative application is built into the machine. Humans still defeat machines in terms of sheer creative capacity. Despite this reality, new developments in reinforcement learning, data augmentation, simulation, and stress testing/scenario analytics are moving the goalposts for what is possible for machines. I predict some of the more transformative, end-to-end, enterprise MI deployments in the near future will leverage these algorithmic trends. I look forward to being "proven wrong" by creative developers in other machine-intelligence research spaces that can demonstrate "smarter" learning i.e., applying (and applying creatively) acquired knowledge without loads of stationary data.
The point to take away here is that "machine learning" in a generic sense covers too broad a range of approaches. Discussions need to be anchored in better definitions of what an author/researcher means by a "machine" and "learning."
Finally, my least favorite word in the MI space—artificial. Here the etymology enlightens our discussion. Late 14th century Old French defined artificial as "not natural or spontaneous." By the early 15th century, Middle English speakers refined how the word was used to reflect being "made by man, contrived by human skill and labor." Today, a modern definition of artificial is "Made or produced by human beings rather than occurring naturally, especially as a copy of something natural." So what exactly do we mean when we use the word artificial intelligence? Are we copying natural intelligence (however defined), human intelligence, human organized (enterprise) intelligence? What happens when a machine/algorithm writes its own MI? (This may be neither natural nor produced by a human)
And what part of this intelligence is reflected in the artificial—repetitive processing such as filtering e-mail, skilled efforts such as forecasting interest rates or creating a business strategy? My intent here is to provoke discussion as I am sure many in this space will have answers and strong opinions. Narrowly defined, artificial intelligence can be useful to discuss human-brain replicating algorithms for example); however, the generic AI term typically muddies discussions & planning.
If I had a magic wand, I would eliminate artificial from most of these discussions and replace it with the word/prefix "pseudo-" which means "resembling or imitating." This shift would facilitate clearer discussion as we could label approaches as pseudo-human intelligence, pseudo-groupthink intelligence, pseudo-organizational intelligence, etc. We could also differentiate clearly some of the more innovative research efforts in types of intelligence that may not be a copy of something "natural" e.g., machine-hive intelligence, quantum intelligence (maybe quantum is something that can be defined as "natural"?) You get the point.
Sadly, I doubt I can push against the ubiquitous use of AI & ML. I can only ask that those of us who discuss these topics be specific about what we mean when we discuss a particular type of MI.
The way ahead for MI
According to etymonline.com, artificial intelligence was first coined in 1956, "the science and engineering of making intelligent machines." I am asking we elevate, instead, the term machine intelligence to focus on the science and engineering of using machines to make more intelligent enterprises. We still have the enormous challenge of finding better ways to deploy these varied technologies on an end-to-end basis within financial services enterprises to realize the very real potential to build more productive, more profitable, more efficient, and more resilient FIs.
We will continue this discussion at AMLD and online. I plan to catalyze discussions not only in the area of matching the right project teams & algorithms/methodologies/processes/workflow with well-defined business use cases, but also…
· Data ingestion & curation
· Organization re-engineering
· Hiring, training, retaining, & managing MI (in all its many instantiations) experts
My list is not comprehensive, but it is a good start for solving this enterprise-wide problem common at financial institutions.