Category Archives: Psychology

The Psychology of the Supernatural

This month Kathryn Schulz writes a lovely little essay in the New Yorker about “Fantastic Beasts and How to Rank Them” [1]. As a member of the original International Society of Cryptozoology, as well as life long fan of fictional worlds of all kinds, I enjoyed her summary of recent psychological research on how people think about “impossible” things.

As the title implies, some of this research examines how people reason about imaginary entities and situations. Is a Yeti more or less “impossible” than a vampire? Is levitation more or less impossible than becoming invisible? And so on.

The interesting thing for psychologists is that even though people may agree that something is imaginary and pretty much impossible in the real world, we can not only imagine it, but imagine the world that it exists in.

Of course, imagining the not (yet) real is the heart of creativity of all kinds, so no one should be surprised that people do it. And imagining how a not-yet-real think would really work is the crux of both invention and story telling.

The recent psychological work has worked to tie this imagination with intuitive physics, the unscientific scientific rules that people learn about how the world works. AKA, “commonsense”. For example, objects do not change into other objects. Big objects are generally heavier than little objects. Stuff like this.

Schulz discusses recent experiments that sort through different sets of these rules as they apply to imaginary animals and situations. Essentially, the concept of “impossible” can be broken down into a range of ways that things are impossible. Some things are “impossible” in many, many ways (such as Hollywood vampires or time travel). Others are actually possible, but just not actually factual (such as Yetis or visitors from outer space).

As she notes, at least from what people say in psychology experiments, there is often strong agreement on such decisions. This is very interesting because it offers a glimpse into what and how people learn about the world. These imaginary cases shed light on everyday reasoning about the real and the possible.

As is often the case, psychologists would benefit from walking across the quad to talk with some Anthropologists about this topic.

Schultz hints as some of the cultural variation that can be found in the world. Almost everyone has grown up with tales of ghosts, but the details are different in different traditions. Hollywood has blurred folk cultures with its own super-cultural mish-mash, but Chinese ghosts and vampires are still quite different from English and Transylvanian entities.

These differences are due to the critical role of story telling. Humans like to tell stories,which tie up events, causes, and effects into a coherent narrative. Stories give explanations for random and inexplicable events, and describe the world at a human level.

(Perhaps the key innovation in all of “science” is that it uses a different kind of story, one that isn’t human centered, includes randomness, and is not judged by whether people like the story or not. Stories, yes. But not just any story.)

Many “impossible” animals and situations are known to us through stories, not through experience.  When people are judging Yeti vs.. Dracula they are working from folk tales, not from scientific journals or personal experience. These stories may be based in cognitive illusions (e.g., ideas about disembodied souls) and intuitive physics, but mostly they reflect the motives and anxieties of the society they come from.

Hollywood vampires are scientifically improbable for sure, but some of their features are obviously ideological. Setting aside the evident deep, deep anxieties about the seduction of young women, Hollywood vampires are associated with demonic forces, and are supposed to be allergic to crucifixes and holy water as much as sunlight and silver. These traits is obviously Christian propaganda, painted onto folk tales about revenents. And, by the way, the supposed effects  of a cross on a Vampire is just as plausible or implausible as your beliefs in holy water, crucifixes, and exorcism—which is a whole different psychological question.

Taking verbal reports as indications of folk-science also misses the key point that many such tales have become symbols of specific cultural identity. Endorsing Bigfoot, Biblical literalism, and the everyday influence of demons and angels may be as much about asserting cultural solidarity (or resistance), as a literal claim of truth. This has nothing to do with reason and evidence, and everything to do with personal identity.

I may say that Bigfoot is more likely than Zombies, but maybe that’s reflecting my preference in popular TV shows, and the sub-cultures they reflect.  This belief is social signalling, not pseudo-scientific reasoning about the world.


Finally, I’ll suggest that the psychologists and new anthropologist friends toodle across the quad again, over to the business and law school. Over in that part of campus, the boffins operate in the most dangerous fantasy world of all, one that believes that humans are rational creatures with common sense.

We aren’t. We are fabulists, who believe absurd stories about the world all the time. Any theory that doesn’t take that as an axiom is just plain broken.


  1. Kathryn Schulz, “Fantastic Beasts and How to Rank Them”. The New Yorker.November 6 2017, Conde Nast.

 

Machine Learning Study of Couple Therapy

One of the interesting developments in recent decades has been the deployment of massive computational analyses to observations of human behavior. Even more remarkably, machine learning has proved as good or better at understanding and predicting human behavior as any other method including human judgment (or introspection).

There are many sensors available, which opens the way for all kinds of measurements, including interpersonal behavior. These capabilities have opened a whole new type of social psychology. Alex Pentland called this Social Physics [2].  Md Nasir, and colleagues at USC report on experiments in what they call “Behavioral Signal Processing” [1]. (The same technology is also used for surveillance and persuasion, which are not necessarily in the interests of the subject.)

Of course, the most unique and important human behavior is language. Computers have astonishing abilities to understand speech and written language, despite the absence of any specific “knowledge”, vocal apparatus, or human nervous system. These abilities are simply impossible, according to psychological theory of the 1970s. Yet there they are.

This fall Nasir et al report yet another study, this one measuring the speech of couples in therapy [1]. The machine learning was able to predict outcomes of the therapy as well or better than any other measures, including clinical judgment.

showed that predictions of relationship outcomes obtained directly from vocal acoustics are comparable or superior to those obtained using human-rated behavioral codes as prediction features.“ (p. 1)

The actual study used a large collection of recordings of therapy sessions, but the techniques could be applied to any digitized recording, and likely to live streams of data. One advantage of the prerecorded collection is that it has been hand coded for behavioral features, which can be compared to the machine derived predictions. Also, sufficient time has passed since the collection was recorded to give realistic estimates of long term outcomes.

(One interesting aspect of this study is that the researchers ignored the original comparison conditions. “[O]ur interest … is on predicting relational outcomes independent of treatment received. “ (p. 7))

The speech analysis used common techniques, which are bound to yield a flood of data. In addition to features and statistics for each individual, there were also dyadic measures. The speaking was broken into turn-taking and across the whole interaction, recording various measures of changes.

The machine learning used the rated outcomes to build a classification using various combinations of features. As always, the high dimensional data had to be winnowed down (a process that involves human judgment).

The results are clear: the machine learning essentially tied the human ratings, at least as far as predicting the (human generated) outcome measures.

It is important to note that the machine learning was based on shallow analysis of the speech: loudness, pitch, timing, and so on. No semantic information was included, nor were other modalities such as gestures or facial expressions. The fact that even these relatively trivial features could even tie human judgment is yet another indictment of the unreliability of human intuition about human behavior.

This study is quite suggestive. Perhaps therapists (or self-therapizing individuals, who have fools for clients) might have tools that signal the “state” of a relationship, and help guide the subjects to a better state.

Of course, the models developed in this particular study only predicted the “outcome”. They neither explain the meaning of the variables (just how does the loudness and pacing of speech cause the outcome?), nor even document much in the way of process. If the therapist intervenes to, say, moderate the intensity of their voices, would that have beneficial effects. When and how much would be needed?

Finally, the analysis includes only the two subjects. Shouldn’t the behavior of the therapist be included in the classifier? In principle, the therapist should be doing something, no? Even if it is a placebo effect, it should show up in the machine classifier.

Its early days, but it certainly is exciting to think about creating tools that help people learn to interact with each other in positive ways. And it will be really good to see this technology employed to actually help people, rather than to try to control and manipulate them. (I’m talking to you Facebook, Google, et al.)


  1. Md Nasir, Brian Robert Baucom, Panayiotis Georgiou, and Shrikanth Narayanan, Predicting couple therapy outcomes based on speech acoustic features. PLOS ONE, 12 (9):e0185123, 2017. https://doi.org/10.1371/journal.pone.0185123
  2. Alex Pentland, Social Physics: How Good Ideas Spread – The Lessons From A New Science, New York, The Penguin Press, 2014.

Reconstructing the “First Flower”

If there is anything I love as much as birds, butterflies, and bees, it must be flowers. They are everywhere, and they are beautiful. (They are, after all, all about s*x.)

Flowers and flowering plants emerged over 130 millions years ago, during the Cretaceous period. Flowers emerged during the height of dinosaur times, though it isn’t certain how dinosaurs and flowers may have co-evolved. I like to think that dinosaur predation shaped the evolution of flowering plants, but who knows?

One of the great mysteries of evolution is how flowers began. There are so many flowers, with so many diverse features and designs. There must have been a “first flower”, but what was it like?

This summer the eFLOWER project (A framework for understanding the evolution and diversification of flowers) has published a new study that examines this question [2] .This large group of collaborators augment studies of fossil remains and genetic patterns among living plants with a mathematical model of the evolution of flowers.

The study is based on a large dataset of current and fossil flowers, which has over 13,000 traits. Using information about molecular dating and fossils, they examine possible evolutionary tress. Many, many possible trees.

If I understand the method correctly, the analysis generated possible ‘ancestors’ based on the relationships among current and fossil flowers, and then tested candidates by running thousands and millions of simulated generations of evolution. (I don’t fully understand these computations

This is a large computation!

The result of this heroic effort is a reconstruction of the ‘first flower’, which is bisexual and spirally whirled.

herve.sauquet@u-psud.fr, juerg.schoenenberger@univ Image caption 3D model of the ancestral flower reconstructed by the new study, showing multiple whorls of petal-like organs, in sets of threes.

This finding is interesting because the fossil record shows a radiation of different flowers that share some, but not all these features. In other words, the adaptations would amount to losing features of the ancestral flower.

Our results suggest two different evolutionary pathways for the reduction in number of whorls in early angiosperm evolution.

The authors speculate on possible advantages in such reductions, perhaps supporting increased specialization.

This idea would answer questions about how one kind of flower could evolve into a radically different structure (they all evolve from a common ‘super’ flower). Of course, we now want to know how this first flower might have evolved from ‘pre flower’ plants.

I’m sure this will be a controversial conclusion.

For one thing, it’s a gigantic amount of math, based on data and assumptions that must be examined carefully. I imagine that it will be difficult to independently replicate this computation.

This result calls into question generally held theories based on other methods. Reexamination of the earlier work may or may not yield a new consensus.

It will be interesting to see if additional fossil evidence can be found that documents more of the actual flowers of that period.

It is worth pointing out that this study has generated a visualization of a completely hypothetical flower, which has never existed as far as we know. The wonders of computational science!


  1. eFLOWER. eFLOWER: A framework for understanding the evolution and diversification of flowers. 2017, http://eflower.myspecies.info/.
  2. Hervé Sauquet, Maria von Balthazar, Susana Magallón, James A. Doyle, Peter K. Endress, Emily J. Bailes, Erica Barroso de Morais, Kester Bull-Hereñu, Laetitia Carrive, Marion Chartier, Guillaume Chomicki, Mario Coiro, Raphaël Cornette, Juliana H. L. El Ottra, Cyril Epicoco, Charles S. P. Foster, Florian Jabbour, Agathe Haevermans, Thomas Haevermans, Rebeca Hernández, Stefan A. Little, Stefan Löfstrand, Javier A. Luna, Julien Massoni, Sophie Nadot, Susanne Pamperl, Charlotte Prieu, Elisabeth Reyes, Patrícia dos Santos, Kristel M. Schoonderwoerd, Susanne Sontag, Anaëlle Soulebeau, Yannick Staedler, Georg F. Tschan, Amy Wing-Sze Leung, and Jürg Schönenberger, The ancestral flower of angiosperms and its early diversification. Nature Communications, August 1 2017. https://www.nature.com/articles/ncomms16047

 

Close Reading Apps: Brilliantly Executed BS

One of the maddening things about the contemporary Internet is the vast array of junk apps—hundreds of thousands, if not many millions—that do nothing at all, but look great. Some of them are flat out parodies, some are atrocities, many are just for show (no one will take us seriously if we don’t have our own app). But some are just flat out nonsense, in a pretty package. (I blame my own profession for creating such excellent software development environments.)

The only cure for this plague is careful and public analysis of apps, looking deeply into not only the shiny surface, but the underlying logic and metalogic of the enterprise. This is a sort of “close reading” of software, analogous to what they do over there in the humanities buildings.  Where does the app come from? What does it really do, compared to what they say it does? Whose interests are served?

Today’s example are two apps that pretend to do social psychology: Crystal (“Become a better communicator”) and Knack (“for unlocking the world’s potential”).

[Read Whole Article]

“The technology of touch”

I have frequently blogged about haptics (notably prematurely declaring 2014 “the year of remote haptics”), which is certainly a coming thing, though I don’t think anyone really knows what to do with it yet.

A recent BBC report  “From yoga pants to smart shoes: The technology of touch”  brought my attention to a new product from down under, “Nadi X”, “fitness tights designed to correct your form”. Evidently, these yoga pants are programmed to monitor your pose, and offer subtle guidance toward ideal position via vibrations in the “smart pants”.

(I can’t help but recall a very early study on activity tracking, with the enchanting title, “What shall we teach our pants?” [2]  Apparently, this year the answer is, “yoga”.)

Source: Wearable Experiments Inc.
Source: Wearable Experiments Inc.

It’s not totally clear how this works, but it is easy to imagine that the garment can detect your pose, compute corrections, and issue guidance in the form of vibrations from the garment. Given the static nature of yoga, detecting and training for the target pose will probably work, at least for beginners. I’d be surprised if even moderately experienced practitioners would find this much help, because I don’t know just how refined the sensing and feedback really will be.  (I’m prepared to be surprised, should they choose to publish solid evidence about how well this actually works.)

Beyond the “surface” use as a tutor, the company suggests a deeper effect: it may be that this clothing not only guides posture but can create “a deeper connection with yourself”. I would interpret this idea to mean, at least in part, that the active garment can promote self-awareness, especially awareness of your body.

I wonder about this claim. For one thing, there will certainly be individual differences in perception and experience. Some people will get more out of a few tickles in their trousers than others do. Other people may be distracted or pulled away from sensing their body by the awareness of their garment groping them.

Inevitably, touch is sensual, and quickly leads to, well, sex. I’m too old not to be creeped out by the idea of my clothing actively touching me, especially under computer control. Even worse, when the computer (your phone) is connected to the Internet, so we can remotely touch each other via the Internet.

Indeed, the same company that created Nadi X created a product called “fundawear” which they say is, “the future of foreplay” (as of 2013).  Sigh. (This app is probably even more distracting than texting while driving….)

Connecting your underwear to the Internet—what could possibly go wrong? I mean, everything is private on your phone, right?  No one can see, or will ever know. Sure.

I’m pretty sure fundawear will “work”, though I’m less certain of the psychological effects of this kind of “remote intimacy”.  Clearly, this touching is to physical touching like video chat is to face to face. Better than nothing, perhaps, but most people will prefer to be in person.

Looking at the videos, it is apparent that the haptics have pretty limited variations. Only a few areas can buzz you, and the interface is pretty limited, so there are only so many “tunes” you can play. The stimulation will no doubt feel mechanical and repetitive, and probably won’t wear very well. Sex can be many things, but it shouldn’t become boring.

(As a historical note, I’ll point out that, despite their advertising claims, this is scarcely the first time this idea has ever been done. The same basic idea was demonstrated by MIT students no later than 2009 [1], and I’ll bet there have been many variations on this theme.  And the technology is improving rapidly.)


This is a very challenging and interesting area to explore. After following developments for the last decade and more, I remain skeptical about how well any sensor system can really communicate body movement beyond the most trivial aspects of posture.

My own observation is that an interesting source of ideas comes from the intersection of art and wearable technology. In this case, I argue that, if you want to learn about “embodied” computing, you really should work with trained dancers.

For example, you could do far worse than considering the works of Sensei Thecla Schiphorst, a trained computer scientist and dancer, whose experiments are extremely creative and very well documented [4].

One of the interesting points that I have learned from Sensei Thecla and other dancers and choreaographers, is how much of the experience of movement is “inside”, and not easily visible to the computer (or observer). I.e., the “right” movement is defined by how it feels, not by the pose or path of the body. Designing “embodied” systems needs to think “from the inside out”, to quote Schiphorst.

In her work, Schiphorst has explored various “smart garments” which reveal and augment the body and movement of one person, or connect to the body of another person.

Since those early says, these concepts have now appeared in many forms, some interesting, and many not as well thought out as Sensei Thecla.


  1. Keywon Chung, Carnaven Chiu, Xiao Xiao, and Pei-Yu Chi, Stress outsourced: a haptic social network via crowdsourcing, in CHI ’09 Extended Abstracts on Human Factors in Computing Systems. 2009, ACM: Boston, MA, USA. p. 2439-2448.
  2. Kristof Van Laerhoven and Ozan Cakmakci. What shall we teach our pants? In Digest of Papers. Fourth International Symposium on Wearable Computers, 2000, 77-83. http://tmg-trackr.media.mit.edu/publishedmedia/Papers/390-Stress%20OutSourced%20A%20Haptic/Published/PDF
  3. Thecla Schiphorst, soft(n): toward a somaesthetics of touch, in Proceedings of the 27th international conference extended abstracts on Human factors in computing systems. 2009, ACM: Boston, MA, USA. http://www.sfu.ca/~tschipho/softn_alt_chi.pdf
  4. Thecla Henrietta Helena Maria Schiphorst, THE VARIETIES OF USER EXPERIENCE: BRIDGING EMBODIED METHODOLOGIES FROM SOMATICS AND PERFORMANCE TO HUMAN COMPUTER INTERACTION, in Center for Advanced Inquiry in the Integrative Arts (CAiiA). 2009, University of Plymouth: Plymouth. http://www.sfu.ca/~tschipho/PhD/PhD_thesis.html

Bonus video: Sensei Thecla’s ‘soft(n)’ [3].  Exceptionally cool!

 

Cliff Kuang on UX Design for Self-Driving Cars

With news every week about yet more self-driving cars (not to mention Uber’s repeated robotic middle finger to the whole world), it is interesting to read Cliff Kuang’s article in FastCo-Design about “The Secret UX Issues That Will Make (Or Break) Self-Driving Cars” (originally published in February).

The main point of the piece is that for driverless cars to succeed, it is that not getting lost and not killing people isn’t enough. People must want to use them, and, most importantly, feel safe and relaxed.

The goal isn’t to replace the unpleasantness of driving with the unpleasantness of riding in a robot car, it is to replace driving with having a nice ride.  Current efforts fall short on this.

Illustrating this point, Kuang  describes a video of a man trying his self-driving car.

“He hasn’t replaced driving with, say, watching a movie or relaxing—instead, he’s replaced the stress of driving with something worse. He looks at the road, he looks at the wheel, he looks at his hands. He’s scared. And he’s smart to be scared.”

This is a horrible experience, even if the technology works flawlessly.  And there are many such videos on YouTube.

And, as Kuang says, this is a design problem.

In contrast to the YouTube horror shows, he recounts an experience with a self-driving Audi: “The car, by design, was calming me before any worries could surface.

Kuang interviewed Brian Lathrop, who leads Audi’s development effort about how they are designing the experience of operating a care that drives itself. The Audi group is striving to design a self driving car that you can trust.

He boils down the design philosophy to ‘3+1’ things that the human rider/operator needs to know:

  • Who is driving (me or the car)?
  • What is it going to do next?
  • What is the car seeing?
  • When does control transition between me and the car?

The article describes the careful design that tells you what is going on, what is coming soon, and what is possible for you to do. The controls and feedback are prominent and designed to be calming. (They eschew red or green lights, which unconsciously signal ‘right’ or ‘wrong’.) The experience is said to quickly become “boring”—which is actually what they are shooting for.

Another theme in their design is to “retrofit” familiar technology, rather than make up completely new metaphors. For example, one concept uses the familiar steering wheel, pulling it away forward to signal automatic driving, enabling the human to grab the wheel and pull it back to take control. The idea is to feel comfortable and “obvious”.

Lathrop’s training in psychology and experience designing aircraft cockpit controls has taught him to be concerned above all that the human user not be confused about the state of the system.This is what causes air disasters, and will also cause car crashes.

A person operating an automated car needs to clearly understand the state of the car at all times (this is what the 3+1 principles are about). Following this principle, the Audi  has displays that show a diagram of the nearby traffic—to show that the car sees what you see—and indications that a turn is going to happen.

Part of the challenge is to manage human expectations for the technology, both as operators and, in the case of cars, as pedestrians faced with automated vehicles. Expectations are conditioned by a combination of personal experience, by subtle  behavior, and by messages about the capabilities of the system. For example, Tesla’s decision to call the systems ‘Autopilot’ sets expectations far beyond the capability of the current technology.  And a robot car that behaves “politely” enjoys the confidence of pedestrians (rightly or wrongly).

I think it is instructive that this group at Audi has been working for more than four years, patiently learning how to do it right, and how to make the ride “boring”. This contrasts with “the Silicon Valley mindset of just dropping beta tests upon an unsuspecting populace” (and in the case of Uber, shoving them down the throat of the populace). This “beta dropping” is, as Kuang says, “not only naive, but also counterproductive.

My own reaction as I read this article, was “phew!”  It’s a relief that some grown ups are working on the problem.


  1. Cliff Kuang, The Secret UX Issues That Will Make (Or Break) Self-Driving Cars, in Co.design. 2016. https://www.fastcodesign.com/3054330/innovation-by-design/the-secret-ux-issues-that-will-make-or-break-autonomous-cars

 

(PS.  Wouldn’t “Just Dropping Beta” be a good name for a band?)

Robot Wednesday

Google Translate Advances Toward Interlingual Model

When I was young and learning Anthropology and Psycholinguistics, we learned that computer translation, like speech understanding, was essentially impossible, if only because we hadn’t the foggiest notion of how people do these things—so how could we program a computer to do it?

In my lifetime, this certainty has disappeared, though we still haven’t a clue how people do it. First, speech generation and then recognition became extremely reliable, based on probabilistic computational models that successfully mimic human behavior without mimicking humans. It shouldn’t work, but it does.

By the way, similar advances are happening in the analysis of emotions and other non-verbal behavior.  Computers are getting very good at inferring human emotion from sensors, even though we humans have little clue how they do it.

In the past two decades, language translation has become feasible for computers, via various learning processes. Again, these computer translations are very good, but the method has nothing to do with how people understand language (as far as we know). Again, it shouldn’t work, but it does.

This fall researchers at Google have deployed Google’s Neural Machine Translation a system that not only learns to translate between two languages, but learns how to translate between many languages at the same time [1]. A side effect of this process is the ability to, at least sort of, translate between two languages for which there is no sample data (termed “zero shot” translation).

In this sense, the system is learning something about “human language” in general, which linguists and psychologists have been seeking to understand for centuries, without clear success. Wow!

The basic idea is to use large samples of translations (i.e., from humans), to learn enough to translate examples not in the training data. Interestingly, the system works from sentences, i.e., the data is a collection of sentences with corresponding sentences in the second language. Given that a sentence is neither an atomic unit of meaning, nor a complete context for the meaning, it is interesting that the learning works so well from this data. For that matter, there isn’t always a one-to-one translation between sentences in two languages. Theoretically, there isn’t any a priori reason this method should work at all, but it does!

This approach works for one pair of languages, e.g., English to Spanish, but doing it one at a time means that you need N^2 translators for N languages. There are thousands of human languages, and Google currently translates for about 100.

The new system gloms all that into a single model, tagging each example with what the target language is.  (This aspect of the method is trivial!) With these tags, the model learns to translate everything into everything. Cool!

Why wasn’t this done before now? Scale. The combined model and dataset is absurdly large, and takes corresponding computing resources to handle it. The training step for the experiment reported in [1] takes weeks to run on 100 GPUs, which means it would have been impossible even a decade ago.

While the scale is impressive, and the notion of doing many to many learning in a single model is cool, the big headline is that this method seems to (somehow) learn to translate between languages that it has no direct examples of. So, when it learns from a English to Spanish sample, and a Portuguese to English example, the resulting neural model can also do a “transitive” Portuguese to Spanish translation about as well as a model trained for those two languages.

This is cool, and remarkable for “the pleasant fact that zero-shot translation works at all”, it is also “the first demonstration of true multilingual zero-shot translation.” ([1], p. 8)

This unprecedented result leads to pretty basic questions about just what is going on here. How does this “zero shot” translation work? In particular, we wonder if the model is actually learning some kind of abstract, general meta language, and “interlingua”. And if so, how can we understand this interlingua?

The paper offers only the first look at these questions, with some data that offers “hints at a universal interlingua representation”. My own view is that the data suggests that the answer may be complicated, in that the model is likely learning more than one kind of translation. But there is certainly much to study here!

Considering that this sort of machine translation was generally considered to be flat out impossible a few decades ago, and considering that linguists have been fruitlessly searching for an interlingua for centuries, this work is truly remarkable.

As I commented above, it is yet another case where computational methods have achieved performance roughly equivalent to human cognition, even though it is obviously not a model of how human cognition and language works.

When you think about it, this is one of the most remarkable areas of intellectual advance of the early twenty first century. I suspect that, as dinosaurs like me die off, there will be a remarkable synthesis of the immense, laboriously hand-made, legacy of language theory and neurolinguistics, with these empirically derived computational models. The result will be an elegant meta theory of what “human language” actually is, with an understanding of the “design decisions” are incorporated into human nervous systems (and specific computer models), and a concomitant story about how these evolved and relate to our kindred species on Earth.


  1. Melvin Johnson, Mike Schuster, Quoc V. Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, Fernanda Viégas, Martin Wattenberg, Greg Corrado, Macduff Hughes, and Jeffrey Dean, Google’s Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation. Google, 2016. https://arxiv.org/abs/1611.04558
  2. Sam Wong, (2016) Google Translate AI invents its own language to translate with. New Scientist, https://www.newscientist.com/article/2114748-google-translate-ai-invents-its-own-language-to-translate-with/