Emotion Science Keeps Getting More Complicated. Can AI Keep Up?
howwegettonext.com
howwegettonext.com
It ought to be telling that this is a piece about AI written by a historian, citing a psychologist, a historian, and a fictional philosopher - but not a single person with an AI or even computing background. The result is a painfully slow and sloppy retreading of existing ideas. First p-zombies and then the Chinese Room question are sketched out without any reference to that original work, ending with a question-begging assertion that perfect imitation of an emotion is meaningfully different from feeling the emotion.
The first article in the series, grandly titled Silicon Valley Thinks Everyone Feels the Same Six Emotions, is even worse. It critiques Ekman and his work with the TSA, then talks about Alexa, but offers no evidence at all that Alexa or any other Valley project involves Ekman's six-emotion model. (Anyone with a cursory knowledge of machine learning will tell you that it almost certainly doesn't. Reducing inputs down to six artificial categories is not the sort of thing that improves training.)
Given all of that, the parting shot of this piece is especially galling. After indicating total unfamiliarity with AI theory, AI ethics, and post-1970 philosophy in general, the author blithely declares: "if I were contemplating a career in philosophy right now, I’d be thinking about making my central field the ethics of emotion AI. There’s much to be done."
"Empathy" is the name of the reasoning process by which children analyze the experience of their own emotions and behaviors, to draw inferences about the emotional state of others based on observing their behaviors.
But long before that, they are reacting emotionally to the behaviors of others. For example, you can scare a baby.
Along with having a suite of emotions and recognizing cues in others, it's pretty obviously a required piece of learning to read emotion for humans. Below a certain age, it's literally impossible to generate ideas like "the cat doesn't like that", since there's no space between "I like that" and "that is likable". But I don't think I've ever seen somebody describe what kind of structural or chemical changes are correlated with e.g. passing the mirror test.
The self conscious emotions, pride, shame, guilt, envy, remorse, embarrassment, hubris, and many more come from the prefrontal cortex. As it develops over a very long period of time, so do the emotions. I remember that embarrassment doesn't typically develop until you're approximately 7-years-old.
(We have a decent example in autism of how emotion without instinctive mirroring plays out in learning emotional signals. The associations are learned consciously, which works out alright. But I'm not sure there's any precedent for people who have mirroring but nonexistent emotions. Is there something like abulia for lack of emotion? ...is it abulia?)
I'm going to stand by the original statement, though. We have proof that this sort of learning without a backing theory can be done, and no real evidence that having a general theory would make it easier to actually learn the specific associations needed to make predictions. To steal an example from the article: knowing whether the six-emotion model is right or wrong doesn't automatically help us recognize a clenched fist as threat or celebration, and fretting about the impact of psych textbooks on AI without proposing an actual connection isn't particularly compelling.
Replace "think" with "emote" and the insight of the original quote still holds true, I believe. The advanced of artificial intelligence is not built on a bedrock of psychological understanding of human emotion.
For instance, consider the DSM description of borderline personality disorder:
"BPD is a pervasive pattern of instability in interpersonal relationships, self-image, and emotion, as well as marked impulsivity beginning by early adulthood and present in a variety of contexts, as indicated by five (or more) of the following:
- Frantic efforts to avoid real or imagined abandonment
- A pattern of unstable and intense interpersonal relationships characterized by extremes between idealization and devaluation (also known as "splitting")
- Chronic feelings of emptiness
- [...]"
You get the point. You have to meet five or more of the listed criteria. Not four or more. Not six or more. But five or more. There's a lot of ad hoc, informally specified, subjective criteria that probably have no reproducibility if ten different psychologists were all told to independently specify which of the above criteria a given patient satisfies. See the recently published update on the reproducibility crisis: https://www.the-scientist.com/news-opinion/half-the-time--ps....
In general, psychology doesn't seem to be based on many of the lessons acquired in the physical sciences over a very long period of time, and there doesn't seem to be much math involved. Open a textbook on quantum field theory and open a textbook on psychology, and notice the main difference between the two. Data on people is scarce, often biased, improperly sampled, and study results are frequently cherry-picked to promote some sort of predetermined narrative about how the human mind works. The human brain is one of the most complex objects in the known universe, and pretending that we have a modicum of understanding about this object is the type of overconfidence and hubris that will lead to centuries of inadequate mental help for people that need it.
What's the solution? Start collecting data and start trying to quantitatively predict things. I have no doubt that machine learning will fail to predict human behavior, but I also have no doubt that it will fail much less drastically than the current approach. Outcomes need to be strictly measured. In this case, I don't think we are as interested in "explanatory" science as we are "predictive" science; i.e., we don't care as much about why the blackbox model works as long as it improves people's lives. Consider for instance the current system of categorizing personalities: the "Big Five" system (OCEAN). Why is this still a thing? Any sort of primitive GBDT would almost certainly perform better than this, so it's a mystery to me why these crude systems are still being used in the mental health industry in 2018.
Agreed that the DSM is kind of silly in that it draws arbitrary boundaries, but again, it’s pointing at common constellations of characteristics that are absolutely statistically predictive of behavior.
The underlying problem is that we are only just beginning to be able to formally represent complex behaviors in terms of underlying neural activity, and until we can do this fully and properly, there are still useful analyses that can be performed, they’re just “fuzzier”.
One can't start collecting data until one knows what to measure and why. We normally start with problems.
Psychologists have made little progress, as far as I can tell, because they unwittingly chose a hard problem: how does a general problem-solver work? They've actually been trying to solve AGI but without the tools of computer science.
>The human brain is one of the most complex objects in the known universe
Otter hardware seems good enough:
I never said we didn't. Why would anyone collect data without knowing the problem that it will be used for? That's implicit. I don't need to directly state the obvious.
> Otter hardware seems good enough
Ok, here's a technical description of what I meant by "complex object": an object whose output is in theory capable of being represented by a universal Turing machine that is significantly shorter in length than the output itself but is still longer than the output of other objects that humanity has interacted with and is capable of observing. I think you'd be hard pressed to argue that the Kolmogorov complexity of an otter's brain is greater than that of a human's.
It bears emphasis because trying to measure stuff without new theories has been a besetting problem in fields such as psychology and medical research and I think a major reason why many studies can't be reproduced.
>I think you'd be hard pressed to argue that the Kolmogorov complexity of an otter's brain is greater than that of a human's.
Yes. But my point is that we wouldn't need to understand much about how human hardware works if we could only understand how an otter brain works. Beyond that the complexity is due to human culture/memes, which we study in their own right anyhow (in fields such as history, politics, literature).
https://ajp.psychiatryonline.org/doi/abs/10.1176/ajp.161.10....
This is a much better model, but going from research to clinical practice seems to be a very slow process.
> You simply don’t know the field enough to speak intelligently.
If you read the study linked to in the article I cited, the whole field of psychology is having a crisis right now. The Atlantic even titled a recent article "Psychology’s Replication Crisis Is Running Out of Excuses". Your profile page says "academic background in psychology". I wouldn't say having a background in a field that can't replicate its own studies means a whole lot. Am I also not authorized to comment on the field of astrology because I don't have a background in astrology?
> Do not spread false beliefs about a subject you know little about.
Please point out the sentence in my post containing a false belief.
"This is a much better model, but going from research to clinical practice seems to be a very slow process."
You omitted, mistakenly or not, the vital qualifier that the model in question is better not for general personality modeling and prediction, or even psychological pathology in general, but specifically for analyzing the much narrower domain of personality pathology.
In your child post for the submission, you directly imply that psychology does not collect data and try to quantitatively predict things, saying "Start collecting data and start trying to quantitatively predict things." When called out on this error, you post a link to a study that does both of these, and the conclusions of which you misinterpreted.
Yes, the replication crisis is real. Yes, psychology is probably the biggest offender. No, the crisis is nowhere remotely close to only being a major problem in psychology.
Communications and research in QFT and the like are absolutely jammed packed with math because they are the easiest sciences to perform correctly, and are the most amenable to being accurately modeled with simple equations. I am not exaggerating when I say that - remember, a rigorous description of both general relativity and the entire Standard Model fit easily on one standard legal size sheet of paper. You can't even say that about chemistry, let alone even higher level emergent phenomena. Psychology is -hard- to reproduce, not from lack of effort, but because it's fucking hard.
> You seem to be putting words in my mouth. I didn't claim OCEAN wasn't empirically based, nor did I claim that it is used for diagnosis. I said the DSM is used for diagnosis of BPD. I offered OCEAN as an example of the general lack of rigor in the field. Read my post more carefully.
Also, you did claim that it was used for diagnosis. Read your last set of sentences:
> Consider for instance the current system of categorizing personalities: the "Big Five" system (OCEAN). Why is this still a thing? Any sort of primitive GBDT would almost certainly perform better than this, so it's a mystery to me why these crude systems are still being used in the mental health industry in 2018.
I would assume "these" refers to a set of systems inclusive of OCEAN considering it's referred to as the context of the paragraph a few sentences back.
Psychology is real science, that is actually based on making and verifying/refuting precise predictions. Just because applying the fruits of the study is extremely hard, does not mean the study is bogus. That sort of thinking is directly analogous to assuming that computer science is a bogus field just because alot of web programmers literally can't reason about an algorithm even if millions of dollars are on the line.
> In general, psychology doesn't seem to be based on many of the lessons acquired in the physical sciences over a very long period of time, and there doesn't seem to be much math involved.
This is just so wrong. Ten years were spent developing a mathematical system called the Facial Action Coding System: https://en.wikipedia.org/wiki/Facial_Action_Coding_System. This system is used to add together all the muscle movements of the face. When this is done, you can discern from the muscle movements what emotion is being displayed.
Emotion science is known to be a field of psychology that really thrives from science. I can't think of any other field where BA's and BSc's can work together on their research and where MD's can reliably use the work they develop. There's very little fragmentation and the work each group does and contributes to the other is valuable.
Emotions are so deeply tied to the sciences that you can see them occurring in the brain as you give it an MRI. The amygdala activates and registers uniquely for each of the six emotions.
All basic emotions originate from the amygdala. All basic emotions are displayed universally and symmetrically (perhaps with the exception of contempt, but a whole article can be done on that), all basic emotions on the face, all basic emotions are universal, and all basic emotions register as we see other basic emotions.
The last one is interesting. When it comes to clinical psychology, we can actually see patterns in emotion recognition and tie them to mental health disorders. Borderline Personality Disorder scored on par with a baseline, but they selected anger as the emotion when defaulting. Bipolar assumed more emotions were happy during mania periods and sad during depressed periods. Anti-Social saw people who failed to reach par with baseline with either failing to recognize fear or failing to recognize disgust in other individuals.
On top of that, we also know that displays of emotions are important as well. Major Depression saw individuals display a wide array of emotions; almost like they were screaming for someone to see their feelings. And then, psychologists also saw Major Depressive people exhibit minimal emotions with slight movements in the upper face. These people were in a completely different segment and were at a high risk of committing suicide.
We can derive from the muscle movements, the activation of the brain, how these expressions are universal, and how we can mathematically that this science is not "fuzzy". There is a lot of hard work and much consideration from all facets of research that contributes to these results.
Even humans have a lot of problems recognizing full diversity of their own emotional experience, unless trained appropriately.