The Fallacy of Seeing Patterns
blog.clevertap.com
blog.clevertap.com
Instead, replace that phrase with this: "based on brain construction, we humans are predisposed to find patterns in data we encounter". This idea holds up to evolutionary scrutiny, organisms have mechanisms biased for self-protection and finding food and other resources through various forms of pattern recognition. IOW we find patterns we're neurologically capable of finding, particularly in regard to survival and reproduction.
Perhaps it's more accurate to assert we "find patterns" useful in making predictions about our present and future state within the environments we occupy. The fact that patterns may not turn out to be authentic is simply part of a process of refinement of pattern-seeking and improving value as predictors of future states. We may call a pattern an "explanation" but nothing is actually "explained" as shown by the fact we reserve the right to enhance or revise the "pattern", or what we insist it predicts, at any given time.
Edits: grammar and clarity
What I would say the brain is good at is finding patterns in terms of identifying what is appropriate, with a particularly general understanding of "appropriate." There's a huge evolutionary drive for this. It's a bit disjointed to say wolves howl with each other because "brains detect patterns" and somehow find them useful. It's much clearer when we say that "brains encode patterns in terms of appropriateness", and thus wolves howl with each other because their brains know it is appropriate to do so. Just like germ's biology encodes that it's appropriate to wiggle harder in salty water, or how a person knows that certain words are correct in specific situations.
So our brains are more like engines for mapping patterned associations to feelings of appropriateness in context.
Not exactly sure how you define "appropriate" in this context. Like an astronomically complex "neural net", the brain integrates "input" into pattern recognition, so it's a form of classification filtering that allows recognition of phenomena. On the basis of experiences, the classified patterns are bound to probabilistic prediction estimates, in turn informing choice of actions.
Though I know much less about wolf vs. human behavior, if analogous to human speech, howling is an action taken in response to evaluation of current environment, e.g., pattern of other wolf activity, presence of prey, etc. We might infer howling is "appropriate" under some condition, but really it's tautology, because it's equivalent to stating we observe howling under some condition. IOW the latter is a description that's complete and sufficient regarding what we "know".
It might be clearer to rephrase "appropriate" to "what is and what works", though "what works" is probably superfluous to understanding the process. IOW an attempt to "explain" the pattern recognition and response phenomena adds nothing to our knowledge, and in fact adds a level of indirection that tends to obscure the nature of process.
I should let it go at that because it's late and I'm tired. Saying more leads to multi-dimensional consideration of the nature of brain/body operations. If you are really interested in arcane and slippery meta-level or higher-order viewpoints I'm happy to say more, but better when it's earlier in the day and my head is clearer.
I guess I prefer the term appropriate because it more cleanly handles the case when the patterns "don't" match: by indicating that such situations don't exist. There's always a pattern to match, just as there's always some action that's learned (or instinctually) appropriate to any given situation.
If you're trying to see appropriate as "what works", then you're probably more strongly aligning the meaning with a rational process than what I intended, but I don't think you're wrong. Either way, it bears keeping in mind that our day-to-day language isn't really optimized for discussing these kinds of things, so there's bound to be multiple layers of confusion.
After decades-long study of human behavioral phenomena, I'm striving to articulate what I've learned in coherent written form. It's proving difficult to transform a non-linear multi-dimensional model into ordinary English prose that readers can comprehend. So I absolutely agree with your comment about limitations of ability to reduce mental models to common language.
The issues you bring up concerning formalized models that allow mapping behavior to determining factors are indeed of central importance. A model must permit sufficient granularity of analysis, at the same time covering sufficient generality without contradiction of the granular level. The hard part is describing the interactivity of this whole range of "levels", because the immediate and the distant elements are in fact occurring simultaneously and affecting the system under observation in real time. It gets convoluted when we realize the observation itself has effects on the observed behavior.
The problem I have with "appropriate" is the term's ambiguity. OTOH "pattern" implies there's a "match" or there isn't. (I know, patterns can be iffy, but then they're not quite a pattern.) Encountering a situation that's unclear, where no "matched" pattern is evident, immediately arouses alarm. Then we proceed with caution until observing enough that something "familiar" is gleaned, or observe/interact enough to establish a new pattern.
This state of "I don't know" is constantly implicit, patterns never match perfectly, details always vary. Most of the time that's overlooked because we accept a "close enough fit" to established patterns, that is, categorical classification is an abstraction that works adequately most of the time.
For example, often it's good enough to say "that's a tree" without saying what kind of tree. But other times it's important to distinguish a fir from a pine from a hemlock. Patterns are infinitely divisible, ultimately no two trees are identical, at some level of refinement abstractions break down and no longer apply. A thing is no more or less than its actual attributes. Though indispensable for human existence, abstraction is just a tool, pattern recognition is a built-in mechanism of abstraction, best to remember all tools have their limits.
I certainly would never say there's no more to learn, just that defining terms is only a tool for communication, not to be confused with the information we attempt to share. We get confused when we think we are "explaining" phenomena that we observe. In reality, it's less confusing and more informative to simply describe what we observe. Curiously, thoroughly observed phenomena are the things we tend to call self-evident or self-explaining, which suggests an explanation is only an expression of uncertainty about patterns yet to be adequately elucidated.
I'm really not sure what to make of the last line. The goal of analysis should be to produce results that are actionable. In the end it should matter very little how they are obtained as long as they are accurate.
At the end of the day, your point is 100% correct.
Or at least, if you continue to use the metric that diagnosed the problem while trying to address it, you're going to get what you measure. With most groups I've been with, they were relieved to get the metrics good enough to detect anything at all, and don't have the stamina to come up with a more accurate way to determine the same thing.
That's about as actionable of a statement as a CEO telling you he wants you to do whatever "will be the most successful." Well duh :)
Instead it's not clear to me who the target audience is. The phrasing makes it appear to be targeted at analysts and not their business partners. Assuming that to be the case, senior analysts are substantially beyond the level this article is written at (or should be).
Entry level analysts need close supervision to prevent them from making these, and other, mistakes. The examples the author draws (specifically cheese vs. infant mortality and the google flu approximations) don't do a good job at identifying when this issue arises. For the cheese example it's unclear if the phenomenon is real or not (the magnitude of the variation in infant death may actually be significant, if the cheese consumption variation was small then there would be a different story). The author does nothing to help the reader resolve this.
In the Google flu example it's only through hindsight (and colossal failure) that the author identifies the lack of validity in Google's model.
I agree 100% with his point but I don't think the article is providing much value because essentially the author is simply saying: "be aware, this type of problem exists out there..." without providing information necessary to navigate/resolve the problem.
0 matches on this page
Welp, that's pretty sad.
Put differently, you don't learn much by talking, but you learn a lot by listening. You learn a little by writing, but you and your readers learn much more if you read up first. (I am not assuming you wrote this. The themes are general.)
If I wrote a piece that ignored a decades-old, well-known, fundamental result, not only would editors and colleagues slam me for it, I'd be ashamed of it myself. I went back and skimmed a few more of this author's posts and I have to say, they're not of a quality I would suggest to students. If they happen to read this, I hope they'll talk with someone who has a little formal training and revise their work.
Wikipedia references work [0] that indicates Konrad's theory could not be validated empirically.
Do you have any references that validate what you are claiming as a well-known, fundamental result?
Can you give a definition of the theory?
[0] https://www.thieme-connect.com/DOI/DOI?10.1055/s-2007-999113
Nonetheless, let us see whether Konrad's work has been dismantled in more recent studies. A quick trip to PubMed suggests otherwise:
http://www.ncbi.nlm.nih.gov/pmc/articles/PMC2800156/
Note that I did not claim Konrad's result was fundamental. I stated that if I ignored a fundamental result in my writing, I would be pilloried. Konrad's theory of prodromal schizophrenic ideation beginning with this tendency to see patterns where none exist is perhaps of interests to psychiatric historians, but it is not what I would call fundamental. I would claim that if the perfect term for a concept exists and is established, one should use it.
Apophenia, in its now accepted colloquial meaning, is an exceptionally handy shorthand for what the writer describes. Like other swell ideas (Gaussian processes, natural selection) it is so useful that it has accumulated several names (as a parallel, kriging and clonal evolution are alternate). Either would work fine here.
The big issue I have is that it seems impossible to define a general concept of "pattern", such that one could claim it doesn't exist. If such a concept cannot be defined, then it is unclear what the concept actually means. The only thing I can imagine that could possibly fit would be if it meant that a person generates a theory and testable hypothesis that is able to invalidate the theory, and fails to find any supporting evidence when the hypothesis is tested, but still claims that the theory was validated in a way that is demonstrably false. Personally, I don't think psychiatrists are so invested in their patients that they would actually carry out such tests.
The article that you linked primarily recounts a single anecdote where Konrad seems to assume very much about his patient (also, it was unclear about the circumstances; do you know if the anecdote occurred during Nazi rule?). No mention is made of Konrad attempting to verify any if the patient's claims.
What result are you claiming as fundamental? The entire theory seems to be self defeating, as a pattern was suggested that has not been able to be verified.
Konrad did give the phenomenon a catchy name and proposed that an increase in this tendency is an initial step in developing schizophrenia. If someone could actually establish this at a neurogenetic level that would be impressive and fundamental; I'm not aware of anyone doing so. I'd expect it to show up in a CNS journal and NIMH or WT to make a big deal if someone did.
The contrast between epiphany and apopheny is so striking, though, and so relevant to this topic, that it annoys me to no end when it is ignored. At the base of all of statistics is a desire to quantify how much of each is present in an observation, experiment, or cyclic series.
As you probably guessed, I am an applied statistician, not a neuroscientist. (I have serious issues with the way statistics are misused in neuroscience, for whatever that's worth). I do not, and cannot, claim that Konrad's theory is fundamental to that field. I do claim that anyone attempting to explain statistical reasoning to a lay public ought to internalize the contrast he proposed. Its setting as a proposed turn towards insanity is just a happy historical note.
I don't think pattern recognition drives most gamblers. There are all kinds of other benefits, perceived or real, that are not accounted for in a purely monetary payoff grid.
I don't know how you can generate random noise. I assume you are using a standard, pixelated display to read this message. Even if a random process was choosing what to display on that screen, there are only a finite number of configurations. Exactly what you are viewing now could be recreated by such a random process.
The problem that hasn't been addressed yet is that "pattern" is not well-defined. If a random display shows a horizontal line pattern, it is still a pattern by some definition (and you would have no way to distinguish it from a "intentionally patterned" display that has the same configuration).
Since you're not going to get proof one way or another, all a well designed and experiment can do is give you evidence. This happens to be more valuable than just about anything else that science has come up with, but it isn't proof.
Which is why the gold standard for a result is replication in a large sample. I could have this very page generated by convolving couple of high entropy random streams. Is it likely to happen repeatedly? Not if the generator is any good. Same principle for randomized trials. You can end up with unbalanced arms (I'm proofing a manuscript where we had exactly this problem). But it's unlikely that they'll be consistently unbalanced across trials with sufficient sample sizes.
Outside of math, there is (almost?) no absolute proof. The weight of the evidence is all we have.
Why use two pages when one word will do?
I looked at the author's other work. It is quite poor. This is clickbait.