Machines + Society #12: Tools for our brains; Social sphere packing; Fox Hunts

A newsletter by Mako Shen

Comment 4Share

machines + society

Mako Shen | Jul 31, 2020

Let’s Take Our Brains More Seriously When Learning

For those seeing this via email, click here to read the latest version of the article.

[William Hogarth's 1736 engraving, Scholars at a Lecture, via Wikipedia]

Books are horrible at teaching. So are lectures. So are documentaries. If to teach means to impart lasting knowledge, then our current educational media are lousy teachers. We encounter thousands of concepts a year, yet the vast majority fades from our memory into the ether, never to be held again. We forget.

This wouldn’t be such a big problem if it weren’t for our denial. We overestimate the stability of our memories, and think we understand concepts better than we do. As our smartphones enable us to access more information, we become convinced that we understand more and more about the world. Yet not only do we understand less and less, we also become increasingly shaped and enfeebled by our technology. (We no longer have to be able to distinguish between edible and poisonous plants. We don’t even need to know exactly how to drive across the city without our phones).

I. Does it have to be like this? Why don’t we have more media experimentation?

As Andy Matuschak points out, "[b]ooks don't work for the same reason that lectures don't; neither medium has any explicit theory of how people actually learn things". Implicit in the medium of books and lectures is the idea of transmissionism— just as words can be transcribed from a screen to a notepad, the teacher’s words are transmitted more or less directly into a student’s mind. This has long been debunked. People have idiosyncratic ways of grasping new concepts; the way that I learn geometry is more visual than my friend, who prefers symbolic equations. We also forget information at an exponentially decaying rate; more than half of what we process is gone within the first 20 minutes. This means that we need several prompts to make persistent memories. (For more on memory failure modes, see Schachter’s Seven Sins of Memory).

The Ebbinghaus Forgetting Curve:

The equation approximation to the forgetting curve:

Both the vast increase in the amount of content (thanks to the internet) and its increasingly bite-sized nature (thanks to social media) can’t be helping our retention either. More information means each new piece of knowledge is less likely to be remembered. A decreased tolerance for mental discomfort means complex concepts are given less airplay. The internet has shortened the lifespan of cultural ideas.

The question around the lack of media experimentation is closely tied to another: why don’t knowledge workers take learning more seriously? Olympic athletes often have training schedules that specify both the exact exercises and the order that they should be performed. Nutrition, recovery, and sleep times are also carefully regimented— the Olympic gymnast Simon Biles has an elaborate recovery routine involving Epsom salt baths, foam rolling, and compression boots. I’ve been fortunate to have surrounded myself with several brilliant people in the labs I’ve worked in around Cambridge, and yet I can report that basically nobody has a schedule resembling the same rigor as that of a professional athlete. This is not to say that researchers don’t work hard— they absolutely do. It’s that information is gathered inconsistently, and that there is very rarely an organized process for testing recall/understanding of new facts. One exception is a very dedicated medical student that I know, who, owing to the intense memorization required in medical school, has a regimen using spaced repetition software. Yet still I predict that this habit will fall by the wayside after he gets his medical degree.

Why don’t we researchers do this? We know we forget, and spaced repetition is proven to significantly boost recall. Recall is required for complex understanding. Are the incentives missing?

Here are some ideas:

  • We are bad at assessing skill. This means that there is only a fairly weak direct incentive to become highly knowledgeable in your field since you can ‘make it’ by learning just enough to appear proficient.

  • The tools are not well-known/good/easy to use. Anki, for all its popularity among medical students, has decks that are fairly clunky.

II. Where can we go from here?

It seems to me that the tools we currently have access to barely scratch the surface of what’s possible. We have learned so much about how the brain works in the last 50 years; if we think carefully about how to use these insights, we can build tools that actually empower, rather than enfeeble us. Some useful facts about the brain:

  • Our associative memory: how can we use the idiosyncratic links between concepts to organize our thought? We don't think in terms of hierarchies or sequential memories so much as associations.

  • Our predictable forgetfulness: as demonstrated by Ebbinghaus, the way we forget is fairly consistent. Can we make tools to i) better predict the things we will forget, and ii) prompt us to remember as efficiently as possible?

  • Our analogous thinking: we often make models of the world by pulling metaphors in other domains— for instance, “the brain is like a rider (the conscious, planning system two) atop an elephant (the subconscious, instinctive system one)— the rider can sometimes sway the elephant but his/her control is precarious”. Can we make tools that make analogizing easier? What about a database of useful analogies that we can draw upon when learning a new concept?

  • Our emotional thinking: when an idea or situation is emotional, it stays with us far longer than otherwise. Can we create systems that allow us to ‘emotionalize’ important concepts or facts so that they stay with us for longer?

  • Visual/spatial memory: our memory is, in a deep sense, spatial (the 2014 Nobel prize in Physiology or Medicine was awarded for the discovery that certain cells in the human hippocampus encode a neural representation of Euclidean space) What about a tool that helps us create a more explicit spatial representation for concepts? Cicero talked about the method of loci (aka ‘memory palace’) for remembering intricate details. How can we create a better modern version?

These characteristics are far from exhausting and are certainly not mutually exclusive.

Now let me point to two experiments that really excite me.

The first is called Quantum Country. The authors, Michael Nielsen and Andy Matuschak, call it a “a prototype for a new type of mnemonic medium”. The key feature is that they have a spaced repetition system (SRS) built-in. You sign up with your email address and, as you read through the essay, you get prompts that check your understanding:

Afterwards, you’ll get an email in a few days with a link to refresh your memory on some of the concepts that you’d gotten wrong. This is a really powerful idea because it lowers the friction of creating a memory prompt by outsourcing the question to the writer of the article and integrates the prompt into the essay itself. What if this became widespread enough on education websites that we could have a personal ‘deck’ of questions that we could review and answer? We could retain so much more of what we read online.

The second is a personal knowledge management tool called Roam Research. It’s difficult to explain the power of this tool in a paragraph, but suffice it to say Roam embraces the associative nature of our thinking by making page creation really really easy.

For instance, if I’m writing about AI governance in China, I can embed a series of other concepts into my paragraph by adding “[[ ]]”around an idea. I can embed an idea like so: “[[Chinese researchers are taking AI Governance fairly seriously]]: it was the main theme of their World AI Conference in 2019…”

This easy embedding means that a natural graph of ideas gets created as you write.

I can now organize and review complex networks of ideas for everything I’ve ever written about. This is not just a different way of taking notes, but a new tool for thinking. It’s difficult to convey just how powerful I think this could be. If you want to read more, I suggest Nat Eliason’s post. (Note that Obsidian and Remnote are two excellent alternatives with slight differences).

What is common to both of these tools is the exponential return to increased effort. Spending 20% more time studying will typically yield a 20% or smaller increase in your test score. In the case of spaced repetition, however, 20% more time would mean over 10 times the retention (12.6 times, if we believe use the Ebbinghaus equation). Similarly, a 20% increase in the effort put into annotation in Roam yields a graph that is qualitatively far richer (this is a more fuzzy claim, but if you begin using Roam, I believe this will be apparent).

Alan Kay said that the best way to predict the future is to invent it. If digital technology is going to be with us for the long term, the future is probably best if our technology empowers, rather than displaces, us.

I’m going to do something different this essay. Here are some prompts to test your recall/understanding:

  • What, specifically, is the problem with how we consume media now if we want to learn?

  • What are some dimensions to consider if we want machines to augment our ability to remember?

(Next month, I’ll ask again.)

Further resources:

📰 Assorted Links 📰

What I'm Noticing

This was a bad month for U.S.-China relations. Notably, the U.S. State Department called for the Chinese Houston embassy to be shut down. The State Department alleged that this embassy was part of an elaborate espionage and influence operation. The details are scary. Chris Wray, the director of the FBI, has called attention to the CCP blackmailing Chinese-born dissenters (many of which are green card holders or U.S. citizens) to return to China. It’s called the “Fox Hunt program” (etymology unclear), and Wray says that some of the victims families in China have been arrested “for leverage,” “'We’ve now reached a point where the FBI is opening a new China-related counterintelligence case about every 10 hours.”

On top of this, the Taiwanese foreign minister is increasingly worried about a military clash with the Chinese army. “China has conducted an “unprecedented” number of sea and air drills around Taiwan in 2020, with the pace rising to nearly once every day since June.” Note that he doesn’t think an invasion of the Taiwan mainland is likely, but rather one of Taiwan’s outlying islands.

I like these sociological studies about large level group dynamics:

Each study has a distinct methodology for studying societal-level effects. I pay attention to the developments of these methods because it dictates the types of conclusions we can draw about society. I may write a post about this in the future.

There is such a thing as a ‘hot-hand’ in specific cases. The chance of making a three-pointer after two previous successes is somewhat greater than taking a cold three-pointer shot. This doesn’t, however, debunk the hot-hand fallacy. See the Data Colada blog post on this. Warning: highly unintuitive probability explanations included.

In praise of negativity. Why being constructively critical is one of the most useful things you can do for collective thought. In short, because we are far better at detecting falsehoods in others than ourselves, a culture of constructive social criticism is far more useful to get to the truth. [Crooked Timber]

Escaping Paternalism Book Club: Part 1. Some good discussion about the shortcomings of the approach that Thaler and Sunstein (and accordingly, a large number of behavioral economists) use to rationalize a number of government interventions. Thaler and Sunstein claim that their interventions should support people’s decisions “[as] if they had complete information, unlimited cognitive abilities, and no lack of willpower.” (Sunstein and Thaler 2003, 1162) They reject that actions reveal preferences, yet fail to provide a means of discerning people’s actual preferences. This is a really hard problem, and is something people in the AI safety community are very aware of. It is a major component of the value alignment problem. It would be interesting to see behavioral economists engage more with what the decision theorists/computer scientists have thought about (and vice versa, though it seems that AI safety people are more aware of the libertarian paternalist literature). [Econlib]

There has been a slow rollback of COVID benefits in the corporate world. “The FCC’s pledge for companies to ensure customers don’t lose their broadband or telephone connections during the pandemic also ended June 30 — now it’s up to individual service providers to decide.” [source: Vox]

Miscellaneous:

  • Why Humans Totally Freak Out When They Get Lost. “ When the aviator Francis Chichester was teaching navigation to RAF pilots during the Second World War, two of his students went missing during an exercise. Chichester searched for them for days in his light aircraft in the Welsh hills, without success. Three months later, he heard that they were prisoners of war: They had misread their compass and flown 180 degrees in the wrong direction, traveling southeast instead of northwest, and had crossed the English Channel thinking it was the Bristol Channel.” [Wired]

  • Social distancing can be formulated as a sphere packing problem. [Quanta]

  • “While Modern English has a two-form system of yes and no for affirmatives and negatives, earlier forms of English had a four-form system, comprising the words yea, nay, yes, and no. Yes contradicts a negatively formulated question, No affirms it; Yea affirms a positively formulated question, Nay contradicts it.

    Will they not go? — Yes, they will.

    Will they not go? — No, they will not.

    Will they go? — Yea, they will.

    Will they go? — Nay, they will not.” [Wikipedia]

Words of the month

Musth: Musth or must is a periodic condition in bull elephants characterized by highly aggressive behavior and accompanied by a large rise in reproductive hormones. Testosterone levels in an elephant in musth can be on average 60 times greater than in the same elephant at other times. [Wikipedia]

Nobodaddy: (A disrespectful name for) God, especially when regarded anthropomorphically. Also (in extended use): a person no longer held in high esteem. [Lexico] More on William Blake’s neologism.

  🎧 Music 🎧

Alash Ensemble for some Mongolian throat singing.

Insensatez: A Mulher Que Fez. Brazilian. Very catchy.

Magnolia — Jorge Ben Jor. O que eu quero mais? (What more do I want?)

Comment 4Share
← PreviousNext →