Less Wrong is a community blog devoted to refining the art of human rationality. Please visit our About page for more information.

## Model Combination and Adjustment

49 17 July 2013 08:31PM

The debate on the proper use of inside and outside views has raged for some time now. I suggest a way forward, building on a family of methods commonly used in statistics and machine learning to address this issue — an approach I'll call "model combination and adjustment."

### Inside and outside views: a quick review

1. There are two ways you might predict outcomes for a phenomenon. If you make your predictions using a detailed visualization of how something works, you're using an inside view. If instead you ignore the details of how something works, and instead make your predictions by assuming that a phenomenon will behave roughly like other similar phenomena, you're using an outside view (also called reference class forecasting).

Inside view examples:

• "When I break the project into steps and visualize how long each step will take, it looks like the project will take 6 weeks"
• "When I combine what I know of physics and computation, it looks like the serial speed formulation of Moore's Law will break down around 2005, because we haven't been able to scale down energy-use-per-computation as quickly as we've scaled up computations per second, which means the serial speed formulation of Moore's Law will run into roadblocks from energy consumption and heat dissipation somewhere around 2005."

Outside view examples:

• "I'm going to ignore the details of this project, and instead compare my project to similar projects. Other projects like this have taken 3 months, so that's probably about how long my project will take."
• "The serial speed formulation of Moore's Law has held up for several decades, through several different physical architectures, so it'll probably continue to hold through the next shift in physical architectures."

See also chapter 23 in Kahneman (2011); Planning Fallacy; Reference class forecasting. Note that, after several decades of past success, the serial speed formulation of Moore's Law did in fact break down in 2004 for the reasons described (Fuller & Millett 2011).

## Avoid misinterpreting your emotions

67 14 February 2012 11:51PM

A couple of weeks ago, I was suffering from insomnia. Eventually my inability to fall asleep turned into frustration, which then led to feelings of self-doubt about my life in general. Soon I was wondering about whether I would ever amount to anything, whether any of my various projects would ever end up bearing fruit, and so forth. As usual, I quickly became convinced that my life prospects were dim, and that I should stop being ambitious and settle for some boring but safe path while I still had the chance.

Then I realized that there was no reason for me to believe in this, and I stopped thinking that way. I still felt frustrated about not being able to sleep, but I didn't feel miserable about my chances in life. To do otherwise would have been to misinterpret my emotions.

Let me explain what I mean by that. There are two common stereotypes about the role of emotions. The first says that emotions are something irrational, and should be completely disregarded when making decisions. The second says that emotions are basically always right, and one should follow their emotions above all. Psychological research on emotions suggests that the correct answer lies in between: we have emotions for a reason, and we should follow their advice, but not unthinkingly.

The Information Principle says that emotional feelings provide conscious information from unconscious appraisals of situations1. Your brain is constantly appraising the situation you happen to be in. It notes things like a passerby having slightly threatening body language, or conversation with some person being easy and free of misunderstandings. There are countless of such evaluations going on all the time, and you aren't consciously aware of them because you don't need to. Your subconscious mind can handle them just fine on its own. The end result of all those evaluations is packaged into a brief summary, which is the only thing that your conscious mind sees directly. That "executive summary" is what you experience as a particular emotional state. The passerby makes you feel slightly nervous and you avoid her, or your conversational partner feels pleasant to talk with and you begin to like him, even though you don't know why.

To some extent, then, your emotions will guide you to act appropriately in various situations, even when you don't know why you feel the way you do. However, it's important to intepret them correctly. Maybe you meet a new person on a good day and feel good when talking with them. Do you feel good because the person is pleasant to be with, or because the weather is pleasant? In general, emotions are only used as a source of information when their informational value is not called into question2. If you know that you are sad because of something that happened in the morning, and still feel sad when talking to your friend later on, you don't assume that something about your friend is making you feel sad.

People also pay more attention to their feelings when they think them relevant for the question at hand. For example, moods have a larger impact when people are making decisions for themselves rather than others, who may experience things differently. But by default, people tend to assume that their feelings and emotions are "about" whatever it is that they're thinking about at that moment. If they're not given a reason to presume that their emotions are caused by something else than the issue at hand, they don't.2

## The Outside View Of Human Complexity

14 08 October 2011 06:12PM

One common question: how complex is some aspect of the human body? In addition to directly evaluating the available evidence for that aspect, one fruitful tactic in making this kind of prediction is to analyze past predictions about similar phenomena and assume that the outcome will be similar. This is called reference class forecasting, and is often referred to on this site as "taking the outside view."

First, how do we define complexity? Loosely, I will consider a more complex situation to be one with more components, either in total number or type, which allows for more degrees of freedom in the system considered. Using this loose definition for now, how do our predictions about human complexity tend to fare?

Point: Predictions about concrete things have tended to overestimate our complexity

Once we know about their theoretical existence of phenomenon but before they are systematically measured, our predictions about measurable traits of the human body tend to err on the side of being more complex (i.e., more extensive or variable) than reality.

1) Although scholars throughout history have tended to think that human brains must be vastly differently from those of other animals, on the molecular and cellular level there have turned out to be few differences. As Eric Kandel relates in his autobiography (p. 236), "because human mental processes have long been thought to be unique, some early students of the brain expected to find many new classes of proteins lurking in our gray matter. Instead, science has found surprisingly few proteins that are truly unique to the human brain and no signaling systems that are unique to it."

2) There turned out to be fewer protein-coding genes in human body than most people expected. We have data on this by way of an informal betting market in the early 2000's, described here (\$) and here (OA). The predictions ranged from 26,000 - 150,000, and that lower bound prediction won, even though it probably wasn't low enough! As of 2008, the predicted number by Ensembl was in the 23,000s. (As an aside, humans don't have the largest genome in terms of number of nucleotides either, by far. That title currently belongs to the canopy plant, pictured below (thanks to kodamatic for the photo, and to Pellicer et al. for the sequencing effort).)

3) Intro neuro texts (including one co-written by the aforementioned Kandel) claim that there are 10-fold (or more) more glia cells than neurons in the human brain. Since glia play crucial support roles and can even propagate info signals, this is not a trivial claim and would vastly increase the processing power of the brain. But when it has actually been measured, the ratio of glial to neural cells is actually around one to one in most species, including humans (see here and here).

Counterpoint: Categories we use to explain the function of our bodies have tended to be more arbitrary than we recognize

1) One active area of research is in determining whether the distinguishing characteristics between what we consider cell "types" are more quantitative or qualitative (i.e., degree rather than form). Consider, for example, the continuum between the "classical" m1 and "alternative" m2 macrophages, which contributes whether those immune cells will be pro- or anti-tumor. Or consider the gradient of pluripotency in stem cells. If cell types are on a spectrum, depending upon the sort of transcripts or proteins they contain at any given moment, that suggests that they may be able to have more different sorts of interactions at different points in time.

2) Although we found fewer human genes than most geneticists expected, components of genes (exons) have been found to be able to combine in many ways, a phenomenon called alternative splicing. One article (here) found that of genes with multiple exons, more than 90% are alternatively spliced. Specifically, these researchers found ~67,000 alternatively spliced transcripts from ~20,000 genes. Since these alternatively spliced genes have different nucleic acid sequences, they could (and probably do) have quite different functions.

3) The chromatin state of a given portion of the genome, i.e. where it falls on the spectrum of euchromatic vs heterochromatic, seems to have the ability to explain a large percentage of a variance in whether or not that gene is expressed. For example, one study (here) shows a strikingly high correlation between the ability of one transcription factor to bind to DNA and the chromatin state of that region of DNA (check figure 3). The fact that these chromatin states can be transmitted between generations via germ cells is also a fascinating finding that has implications which increase the complexity of human biology as compared to the "static DNA" model.

Synthesis: When to expect more or less complexity

The above is far from systematic, but I think it portrays the trends. The known unknowns have tended to end up lower in complexity than we've predicted. But unknown unknowns continue to blindside us, unabated, adding to the total complexity of the human body.

Why do we tend to over-estimate the complexity of known unknowns in the human body? People who study biological processes want to find more "degrees of freedom" in their systems, so that the phenomenon they're studying can have more explanatory power. The standard reason for this is that they want their results to have an impact in preventing or curing diseases, while the cynical ("Hansonian") reason is that they want to attract more status and funding. The real answer is probably a mix of both, but either way, the result is that we tend to over-estimate the complexity of the known unknowns.

Why does it take so long to recognize the vast number of unknown unknowns? I think the best explanation for this is the standard, "Kuhnian" one, that shifting a paradigm is difficult. Adding an entirely new facet to any established scientific discipline requires slow-moving institutional support, and human biology is no exception. Look, for example, at the history of neurogenesis. Another explanation is technological, that we just don't have the capacity to observe certain things until we reach a given level of engineering success. We could not have known about histone-based epigenetics until we had the capacity to visualize cells at the level of electron microscopy (see pdf).

The next time someone uses an argument like "the human body is so complex," try to notice whether they are referring to a prediction about the way that the human body and biology work in general, or one particular aspect of the human body. If they're referring to the general issue, at scales from the atomic to the molecular to the tissue level, they're right: there's loads we don't understand and probably lots of important stuff we don't even know about. But if they're referring to a particular as-of-yet unmeasurable aspect of the human body, past history suggests that that particular phenomenon is likely to be less complex than you might guess.

References

Kandel, E. In Search of Memory: The Emergence of a New Science of Mind. amazon

Pennisi, A. 2003 Low Number Wins the GeneSweep Pool. abstract.

Human Genome Information Project. 2008 How Many Genes Are in the Human Genome?. link

Pellicer J. 2010 The largest eukaryotic genome of them all? abstract. doi: 10.1111/j.1095-8339.2010.01072.x

Kandel E, et al. Principles of Neural Science. amazon

Azevedo FA, et al. 2009 Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. pubmed

Ma J, et al. 2010 The M1 form of tumor-associated macrophages in non-small cell lung cancer is positively associated with survival time. doi:10.1186/1471-2407-10-112

Hough SR, Laslett AL, Grimmond SB, Kolle G, Pera MF (2009) A Continuum of Cell States Spans Pluripotency and Lineage Commitment in Human Embryonic Stem Cells. PLoS ONE 4(11): e7708. doi:10.1371/journal.pone.0007708

Toung JM. 2011 RNA-sequence analysis of human B-cells. abstract. doi:10.1101/gr.116335.110

John S, et al. 2011 Chromatin accessibility pre-determines glucocorticoid receptor binding patterns. doi:10.1038/ng.759.

Olins DE and Olins AL. 2003 Chromatin history: our view from the bridge, pdf

Wheeler A. A Brief History and Timeline: Adult mammalian neurogenesis. link.

## "Outside View!" as Conversation-Halter

50 24 February 2010 05:53AM

Followup toThe Outside View's Domain, Conversation Halters
Reply toReference class of the unclassreferenceable

In "conversation halters", I pointed out a number of arguments which are particularly pernicious, not just because of their inherent flaws, but because they attempt to chop off further debate - an "argument stops here!" traffic sign, with some implicit penalty (at least in the mind of the speaker) for trying to continue further.

This is not the right traffic signal to send, unless the state of knowledge is such as to make an actual halt a good idea.  Maybe if you've got a replicable, replicated series of experiments that squarely target the issue and settle it with strong significance and large effect sizes (or great power and null effects), you could say, "Now we know."  Or if the other is blatantly privileging the hypothesis - starting with something improbable, and offering no positive evidence to believe it - then it may be time to throw up hands and walk away.  (Privileging the hypothesis is the state people tend to be driven to, when they start with a bad idea and then witness the defeat of all the positive arguments they thought they had.)  Or you could simply run out of time, but then you just say, "I'm out of time", not "here the gathering of arguments should end."

But there's also another justification for ending argument-gathering that has recently seen some advocacy on Less Wrong.

An experimental group of subjects were asked to describe highly specific plans for their Christmas shopping:  Where, when, and how.  On average, this group expected to finish shopping more than a week before Christmas.  Another group was simply asked when they expected to finish their Christmas shopping, with an average response of 4 days.  Both groups finished an average of 3 days before Christmas.  Similarly, Japanese students who expected to finish their essays 10 days before deadline, actually finished 1 day before deadline; and when asked when they had previously completed similar tasks, replied, "1 day before deadline."  (See this post.)

Those and similar experiments seem to show us a class of cases where you can do better by asking a certain specific question and then halting:  Namely, the students could have produced better estimates by asking themselves "When did I finish last time?" and then ceasing to consider further arguments, without trying to take into account the specifics of where, when, and how they expected to do better than last time.

From this we learn, allegedly, that "the 'outside view' is better than the 'inside view'"; from which it follows that when you're faced with a difficult problem, you should find a reference class of similar cases, use that as your estimate, and deliberately not take into account any arguments about specifics.  But this generalization, I fear, is somewhat more questionable...

## Conversation Halters

38 20 February 2010 03:00PM

Related toLogical Rudeness, Semantic Stopsigns

While working on my book, I found in passing that I'd developed a list of what I started out calling "stonewalls", but have since decided to refer to as "conversation halters".  These tactics of argument are distinguished by their being attempts to cut off the flow of debate - which is rarely the wisest way to think, and should certainly rate an alarm bell.

Here's my assembled list, on which I shall expand shortly:

• Appeal to permanent unknowability;
• Appeal to humility;
• Appeal to egalitarianism;
• Appeal to common guilt;
• Appeal to inner privacy;
• Appeal to personal freedom;
• Appeal to arbitrariness;
• Appeal to inescapable assumptions.
• Appeal to unquestionable authority;
• Appeal to absolute certainty.

Now all of these might seem like dodgy moves, some dodgier than others.  But they become dodgier still when you take a step back, feel the flow of debate, observe the cognitive traffic signals, and view these as attempts to cut off the flow of further debate.

## In defense of the outside view

14 15 January 2010 11:01AM

I think some of our recent arguments against applying the outside view are wrong.

1. In response to taw's post, Eliezer paints the outside view argument against the Singularity thus:

...because experiments show that people could do better at predicting how long it will take them to do their Christmas shopping by asking "How long did it take last time?" instead of trying to visualize the details.

This is an unfair representation. One of the poster-child cases for the outside view (mentioned by Eliezer, no less!) dealt with students trying to estimate completion times for their academic projects. And what is AGI if not a research project? One might say AGI is too large for the analogy to work, but outside view helpfully tells us that large projects aren't any more immune to failures and schedule overruns :-)

2. In response to my comment claiming that Dennett didn't solve the problem of consciousness "because philosophers don't solve problems", ciphergoth writes:

This "outside view abuse" is getting a little extreme. Next it will tell you that Barack Obama isn't President, because people don't become President.

## Reference class of the unclassreferenceable

25 08 January 2010 04:13AM

One of the most useful techniques of rationality is taking the outside view, also known as reference class forecasting. Instead of thinking too hard about particulars of a given situation and taking a guess which will invariably turned out to be highly biased, one looks at outcomes of situations which are similar in some essential way.

Figuring out correct reference class might sometimes be difficult, but even then it's far more reliable than trying to guess while ignoring the evidence of similar cases. Now in some situations we have precise enough data that inside view might give correct answer - but for almost all such cases I'd expect outside view to be as usable and not far away in correctness.

Something that keeps puzzling me is persistence of certain beliefs on lesswrong. Like belief in effectiveness of cryonics - reference class of things promising eternal (or very long) life is huge and has consistent 0% success rate. Reference class of predictions based on technology which isn't even remotely here has perhaps non-zero but still ridiculously tiny success rate. I cannot think of any reference class in which cryonics does well. Likewise belief in singularity - reference class of beliefs in coming of a new world, be it good or evil, is huge and with consistent 0% success rate. Reference class of beliefs in almost omnipotent good or evil beings has consistent 0% success rate.

And many fellow rationalists not only believe that chances of cryonics or singularity or AI are far from negligible levels indicated by the outside view, they consider them highly likely or even nearly certain!

There are a few ways how this situation can be resolved:

• Biting the outside view bullet like me, and assigning very low probability to them.
• Finding a convincing reference class in which cryonics, singularity, superhuman AI etc. are highly probable - I invite you to try in comments, but I doubt this will lead anywhere.
• Or is there a class of situations for which the outside view is consistently and spectacularly wrong; data is not good enough for precise predictions; and yet we somehow think we can predict them reliably?

How do you reconcile them?

## Kahneman's Planning Anecdote

25 17 September 2007 04:39PM

Followup toPlanning Fallacy

From "Timid Choices and Bold Forecasts: Cognitive Perspective on Risk Taking" by Nobel Laureate Daniel Kahneman and Dan Lovallo, in a discussion on "Inside and Outside Views":

In 1976 one of us (Daniel Kahneman) was involved in a project designed to develop a curriculum for the study of judgment and decision making under uncertainty for high schools in Israel.  When the team had been in operation for about a year, with some significant achievements already to its credit, the discussion at one of the team meetings turned to the question of how long the project would take.  To make the debate more useful, I asked everyone to indicate on a slip of paper their best estimate of the number of months that would be needed to bring the project to a well-defined stage of completion: a complete draft ready for submission to the Ministry of education.  The estimates, including my own, ranged from 18 to 30 months.

At this point I had the idea of turning to one of our members, a distinguished expert in curriculum development, asking him a question phrased about as follows:

"We are surely not the only team to have tried to develop a curriculum where none existed before.  Please try to recall as many such cases as you can.  Think of them as they were in a stage comparable to ours at present.  How long did it take them, from that point, to complete their projects?"

## Planning Fallacy

41 17 September 2007 07:06AM

The Denver International Airport opened 16 months late, at a cost overrun of \$2 billion (I've also seen \$3.1 billion asserted).  The Eurofighter Typhoon, a joint defense project of several European countries, was delivered 54 months late at a cost of £19 billion instead of £7 billion.  The Sydney Opera House may be the most legendary construction overrun of all time, originally estimated to be completed in 1963 for \$7 million, and finally completed in 1973 for \$102 million.

Are these isolated disasters brought to our attention by selective availability?  Are they symptoms of bureaucracy or government incentive failures?  Yes, very probably.  But there's also a corresponding cognitive bias, replicated in experiments with individual planners.

Buehler et. al. (1995) asked their students for estimates of when they (the students) thought they would complete their personal academic projects.  Specifically, the researchers asked for estimated times by which the students thought it was 50%, 75%, and 99% probable their personal projects would be done.  Would you care to guess how many students finished on or before their estimated 50%, 75%, and 99% probability levels?