February 1, 2013

Projective Models: a new explanatory paradigm

Predictive models are well-known, and have been deployed in a number of high-profile technologies. From IBM’s Watson [1] to the development of autonomous aircraft [2], predictive models use a statistical basis for making inferences about future events. Theoretical models of the brain suggest that the brain is a prediction machine, inferring the world from past events [3]. Yet how intelligent can a statistical model be, especially with regard to complex events?  What is needed (and what already exists in the world of futurism) is something called a projective model. Projective models are predictions about the future, but not solely based on past events. Unlike predictive models, projective models are largely based on mental models. They might also be called "blue sky" models [4]. However, they share attributes that might help scientists and futurists build better models of future events.

The success of projective models depends on a tension between uniformity and empirically-driven outliers. Every January 1, lists of predictions for the upcoming year are unveiled. There are elements of uniformity in that historical trends are continued. However, improper predictions of economic collapse and failure to predict deaths and accidents demonstrate the difficulties of properly incorporating outliers. While outlier incorporation is a feature of predictive models, they are a particularly important component of projective models, and a critical component of getting projective statements correct.

I lived in this year. I'm not sure this came to fruition.

But, this could be the way the world works in the 25th century.

In order to make a coherent statement about the future, projective models (as well as predictive models) must assume that trends represent some kind of norm. In projective models, normative parameters are not based on averaging, but on popular heuristics for making a projection. The three most common of these are: 1) assume things stay the same, 2) catastrophe (or comeuppance) is coming soon, and 3) things will decrease/increase linearly. Much like curve-fitting exercises that are part of predictive modeling, projective model heuristics make assumptions about the behavior of the system in question. Likewise, the existence of inherently unpredictable events (such as black swans – [5]) affects projective and predictive models alike.

Three different scenarios for projective models: TOP: things stay the same, MIDDLE: constant increase or decrease over time, BOTTOM: big events change everything.

So the first phenomenon that drives reality away from predictions involve data-driven trends from past that do not necessarily lead to trends in the future. We cannot truly know what the future holds, and so must use extrapolation to make such statements. In the case of economic or other model-based projections (generated using traditional predictive model methodologies), improper assumptions can lead to deceptive results [6]. Sometimes, these extrapolations are interpreted as innumeracy [7]. But since they are imprecise statements, they can sometimes be correct. But what else contributes to incorrect projections?

The other phenomenon that drives reality away from predictions is phase change behavior. Currently, there is a debate as to whether or not the era of economic growth is over in the developed nations. This involves more than just getting the statistical trend correct. It involves getting the transition from one historical phase (growth interspersed with periods of decline) to another phase (dominant periods of flat growth).

Futurism demonstrates the relatively high failure rate of projective models. A recent Paul Krugman blog article [8.1] mentions Herman Kahn's list of predictions for the Year 2000, circa 1967 [9]. Many of these predictions are incorect, but approximately 26% were fulfilled in one way or another. The reasons for this somewhat low success rate and the nature of fulfilling a prediction have led me to come up with three criteria for judging whether or not a prediction is likely to be wrong, fulfilled, or partially fulfilled on a short or moderate historical timescale [10].

What makes a technological prediction fundamentally incorrect?

I will now introduce three reasons why I think future technological projections are often incorrect. These involve historical and physical factors that are often at odds with human intuition and the conception of a mental model:

1) No technological or historical precedent at the time of prediction. If there is no precedent for world peace, why would you predict it to be so fifty years in the future? The amount of development needed not only to make something a reality but to produce it in a replicable fashion (or at the appropriate economy of scale) will roughly determine the amount of time needed to realize the prediction. A related issue involves historical contingency (sometimes also referred to as "lock-in"). It is often easy to predict gains in an existing technological framework [11]. It is much harder to imagine an entirely new paradigm. The movie "Back to the Future, Part II" features an example of this. The movie, partially set in 2015, not only featured flying cars, but a fax machine in every room of the house.

The subject of [8.1, 8.2] is the self-driving car, which was envisioned in the 1990 version of "Total Recall". Now, of course, the self-driving car is becoming a reality. But consider that many of the component technologies that enable the self-driving car (e.g. linear filter, computer vision, GPS) have been around for a few decades. And related varieties of autonomous robot are in the process of changing the economic and social landscape [8.3]. It is incremental developments along that trajectory that have enabled the self-driving car rather than de novo innovation. This is not to say that de novo innovations do not occur or have an effect on the future. Indeed they do, as the rise is innovations surrounding the internet (e.g. social media, online shopping) demonstrate. However, even here, such innovations are dependent on a trajectory of technological advancement and cultural imagination [12].

2) Prediction requires a relatively high energetic threshold (e.g. flying cars). Why don't we have flying cars yet? Or better yet, why are big predators so rare [13]? The answer, or course, involves the energetic requirements involved. If the energetic requirements for a technology are very high (e.g. warp drive), the less likely they will be conceived or developed without an accompanying source of energy. This has, of course, has been the limiting factor for the development of long-range electric cars. How can ecological and predictive models help us understand why projective models can often fail?

In [13], it is argued that ecological constraints (namely, the big predators' energetic footprint) prevent large animals from becoming too numerous. And so it is with the development of technologies that require a high energy density [14]. As much as I think Moore’s Law analogies are severely overused, one is actually appropriate here. According to Moore’s Law, innovations have been able to make the size of transistors decrease by half linearly over time. However, this trend is now being threatened by a fundamental size limitation. Similarly, there exist fundamental energetic limitations to many technologies, and make advances that require high energy requirement improbable.

How big is too big? Can things be scaled up infinitely? Or are there clear energetic limits to realizing certain technologies?

3) Prediction is on a highly complex system (e.g. diagnosis of disease, cyborgs). In general, technologists are either dismissive of complexity or treat it as a quasi-religious mystery. We know from recent unpleasantness in financial markets that complexity can wreak havoc on predictive models. The effects of complexity on projective models are even more problematic. Sometimes, the system that is supposed to be conquered or created to generate the prediction is much more complex than previously assumed.

One example of this comes from the promises made after the initial draft of the human genome was introduced in 2001 [15]. The sequencing work was done faster than expected, and it was assumed that sequence data could provide the necessary information for a curing most diseases in a short period of time. The tenth anniversary was marked with a NYT article [16] contemplating why a lot of the early predictions either never came to fruition or are slow in coming true. What was not taken into account in these early predictions was the sheer complexity of human physiology and its role in disease.

I woke up and discovered complexity! Examples from social (top) and physical (bottom) systems.

4) The Nostradamus Effect: when projective models go bad? The Nostradamus effect can be defined as a prediction that consists of vague, verbal statement can be easily fit to the empirical world. The Nostradamus effect is essentially the opposite of the traditional scientific method. Generally, scientific inquiry proceeds from empirical observations, which result in a theory. The Nostradamus effect uses vague hypotheses that are then matched to future events.

I remember when I was about 10 years old, I was impressed by Nostradamus' predictions, which I was introduced to by a TV presentation by Orson Wells [17]. This inspired a trip to the local library, where I checked out an abridged version of Nostradamus. What I found was underwhelming: his predictions consist of vague axioms which can be applied to any number of scenarios. In the world of predictive models, such a model would be both noisy and prone to false positives.

"Measuring" the future -- a losing battle? Will being more precise in our projections help or hurt their accuracy?

One way to improve the accuracy of projective models is to work towards a semantically-oriented model based on a hybrid Markovian-Bayesian architecture. This would allow us to approximate the current state of the technological innovation as a series of states, but also takes into account conditional information, which accounts for unexpected events. Of course, we would also have to include a number of variables accounting for technological shifts (sudden and gradual), and their underlying causes. Even this would probably not be sufficient for dealing with unpredictable events, and so would also require a stochastic or chaotic component to generate scenarios consistent with the current limitations of future projections.

Even the most elaborate and efficient of models would most likely only make us marginally better than Nostradamus at projecting future trends. Therefore, a way to merge predictive and projective models in a way that would benefit each would be to acquire long historical (or time-series) datasets in the same way we currently acquire (and hype) "big" data [18]. Such an approach will give us both a quantitative and qualitative appreciation of technological evolution that is sometimes missing from many futurist-type predictions.

[1] “The AI Behind Watson”. AAAI technical paper.

[2] Economist article ("This is Your Ground Pilot Speaking") on autonomous aircraft.

[3] Dayan, P. and Abbott, L.F.  Theoretical Neuroscience. MIT Press (1998). AND Hawkins, J. and Blakeslee, S. On Intelligence. Times Books, New York (2003).

[4] For examples of future projections and how they play out in history, see the following features:

1) IEEE Spectrum. Feature: Life in 2030. Podcasts and Videos.

2) Armstrong, S.  Assessing Kurzweil: the results. Less Wrong blog. January 16. (2013)

3) What 2012 Stuff Will Seem Crazy in 2060? The David Pakman Show, YouTube. January 4 (2013).

[5] Taleb, N.N.  The Black Swan: the impact of the highly improbable. Random House, New York (2010).

Visions of the future that involve utopian or dystopian settings are often based on over-interpreting the effects of these large-scale shifts. Especially in terms of the proposed "comeuppance" scenario, a common science fiction trope involves a dystopian outcome from the extreme development of technology (e.g. "1984", "Animal Farm").

[6] See these Wikipedia pages on Statistical Assumptions and Economic Forecasting. And then read these blog posts on economic projections and weather forecasts:

1) Krugman, P.  The Mostly Solved Deficit Problem. Conscience of a Liberal blog. January 10 (2013).

2) Krugman, P.  Future Inequality, according to the CBO. Conscience of a Liberal blog. December 27 (2012).

3) Robson, D.  How good are the Weather Channel's predictions? Short Sharp Science blog. February 6 (2009).

[7] Seife, C.  Proofiness: the dark arts of mathematical deception. Viking Press, New York (2010).

[8] A series of blog posts on the potential of autonomous intelligence and it's effects on the economy and our society in general. This topic is becoming a popular one -- the following are merely a starting point:

1) Krugman, P.  Look Ma, No (Human) Hands. The Conscience of a Liberal blog, January 25 (2013).

2) Thrun, S.  What we're driving at. Official Google blog. October 9 (2010)

3) Kaminska, I.  The Tech Debate Blasts Off. Towards a Leisure Society blog. December 28 (2012).

[9] re-visited and rated courtesy of Leonard Richardson's Crummy blog. Also see his piece on "The Future: a retrospective" (reflections on the the book "Future Stuff").

[10] My definition of short to moderate historical timescales are on the order of 20 to 200 years. This time window can vary based on both context and the current point in time relative to critical sociohistorical events, cultural change, technological revolutions, etc.

[11] Or perhaps not. A good historical yardstick for this is the book "Future Shock" by Alvin Toffler (first published in 1970). Upon reading it in 2013, is it utopian, dystopian, or accurate (or elements of all three)?
Also, here is a link to a documentary on Future Shock narrated by Orson Wells (from 1972).

[12] For more information, please see the following lecture by David Graeber entitled "On Bureaucratic Technologies and the Future as Dream-Time", which discusses the role of existing bureaucracies (social structures) and cultural imagination in technological innovation.

[13] This paraphrases a title of a book by Paul Colinvaux: "Why are Big Fierce Animals Rare?", Princeton University Press, 1979.

[14] A related problem is the improvement of existing technologies, such as more sustainable energy sources for jet aircraft and rockets. Please see this ASME Knowledgebase entry for more information.

[15] International Human Genome Sequencing Consortium  Initial sequencing and analysis of the human genome. Nature, 409, 860-921 (2001) AND Venter, C. et.al  The sequence of the human genome. Science, 291(5507), 1304-1351 (2001).

[16] A series of New York Times articles associated with the 10th anniversary of the draft full human genome sequence:

1) Pollack, A.  Awaiting the Payoff. NYT, June 14  (2010).

2) Wade, N.  A Decade Later, Genetic Map Yields Few New Cures. NYT, June 12  (2010).

3) Editoral: The Genome, 10 Years Later. NYT, June 20  (2010).

[17] "Nostradamous: the man who saw tomorrow". Narrated by Orson Wells (circa 1981). Watch on Vimeo.

[18] Arbesman, S.  Stop Hyping Big Data and Start Paying Attention to ‘Long Data’. Social Dimension blog, January 29.

No comments:

Post a Comment