No, I don't mean that the news media are biased politically.  I mean that authors are biased by the media they use.

I'm learning about support vector machines (SVMs).  There are a lot of books and articles written on SVMs.  There are also a whole lot of video lectures on SVMs at videolectures.net (see "kernel methods").

People go into much greater detail in lectures than in text.  I like to work with text.  I'd like to have a text on SVMs that goes into as much detail as videos on SVMs usually do, and works out the ideas behind the concepts as thoroughly, but no such text exists.  For some reason, giving a 5-hour lecture series in which you describe the motivations, applications, and work out the mathematical details is acceptable; but writing a text of the same level of detail, which might take only 2 hours to read, is not.

Perhaps this is because writers are motivated to keep pagecounts low.  But pagecount no longer matters with electronic articles.  Yet writers still don't want to explain things thoroughly.  They certainly aren't saving their readers any time by leaving out intermediate steps.  A longer article would take less time to read (and possibly less time to write).  Another problem with the pagecount theory is that texts routinely include footnotes and appendices, contributing to the pagecount; yet relegate them to the back of the book, as if embarassed of them, despite the fact that this makes them very difficult to use.

It's especially bad in math, in which writers have a long tradition of deliberately concealing difficult steps and leaving them "as an exercise to the reader".  For some reason it is considered bad form to write out all of the steps in a proof, even if adding one or two lines could save the reader five minutes of puzzling.  I read an electronic article yesterday where the author said, "These two equations are actually equivalent.  Do you see why?"

I think people have adopted a set of cultural biases about what is appropriate in lectures vs. in writing by simply counting observations, without thinking about the systematic sample bias.  Speakers speak the way they've seen other speakers speak, without recollecting that most of those speakers were instructors.  Technical writers, meanwhile, are picking up their cues from authors of textbooks, which are written with the assumption that a person will be on hand to take you through the details; and applying them in situations where no such person will be available.

New Comment
47 comments, sorted by Click to highlight new comments since:
[-]matt150

Audio and video lectures are much less annoying when you play them faster - I find I'm able to listen to most video content at 2.5-3 times normal speed, and most audio at about twice normal speed, both with increased retention (I get less bored, so my mind wanders less).

A couple of time-saving hacks related to speeding up audio or video lectures:

  1. QuickTime is great at tempo shifting video and audio to play it back faster without shifting the pitch ("chipmunking"). Look for QuickTime's A/V controls, and "Playback speed". You have to be sitting at your computer to use this, which brings me to…

  2. If you use a portable media player, PodShifter will shift (speedup without chipmunking) podcast feeds. If your audio isn't in a feed, HuffDuffer will make individual online media files into a feed, to which you can subscribe using PodShifter. PodShifter is free to use. (I have a financial interest in PodShifter, but this comment doesn't feel spammy to me - please downvote it if you disagree.)

The open source Soundtouch library includes soundstretch, which can do audio tempo shifting locally. Audacity can also do local shifting (look for Effects > Tempo shift).

I just listened to 20 minutes of lecture in about 15 minutes of real time, and I had an epiphany: the conventional college lecture is now dead to me. It's still shuffling around a little, but believe me it's dead. Here are some advantages of moving the lectures proper to online video and holding just the discussion/help sessions in person:

  1. The ability to tempo shift. Holy hell this is awesome. I've had terribly boring classes that actually would have been tolerable if only they had been conducted at roughly twice the speed, and even more interesting classes would have been much improved by an extra 50% more speed, preferably along with the ability to continuously vary the amount of tempo shift. The key here is that the lecture goes fast enough to keep your brain turned on; few things are more demoralizing than helplessly wishing you could fall asleep in class.

  2. If the lectures include (say) PDF slides with timing information to synchronize them with the video, things get even better. You can stick in links to web pages, or you can jump around to relevant sections in the video, and generally just have a grand nonlinear time. It's the Wikipedia effect: chasing interesting information around the vast space of human knowledge is a profoundly effective way of learning.

  3. The marginal costs of making the lectures available publicly are tiny. The rising tide of access to education is guaranteed to uplift humanity as a whole, even if Joe Average doesn't notice.

  4. Good lecturers can drive out bad ones. Who hasn't sat in a boring lecture that was taught better by someone else? That's only necessary because class sizes are traditionally limited by the number of warm bodies you can cram into one room.

  5. The learning is available whenever you want it. That flexibility also means that nobody is riding your ass to get you to keep up with your studying, but I'm sure that that service can be provided separately.

By the way, does anybody know a good source of interesting audio lectures with RSS feeds that I can give to PodShifter? I've been looking around iTunes University, and they've got a lot of good stuff, but trying to get an RSS URL from them is like trying to tape weasels to a wall: so irritatingly difficult that you start to wonder why you even tried. The pedagogical revolution needs to get easier, comrades!

Is there software that will compress the empty spaces between words and sentences more than the words and sentences?

I'm pretty sure that's one of the things quicktime does. Probably all of these do, but qt is the only one I've tried.

I think that lectures and presentations are 'friendlier' for at least the following reasons:

  • fewer assumptions are made about the background of the audience in lectures (lectures are often the 1st exposure, aimed at students; papers/textbooks assume more background)
  • the paper/text is expected to be authoritative in a way that a lecture isn't (books/papers contribute to academic reputation, lectures do not)
  • the [primary] purpose of a lecture is to teach, while the [primary] purpose of a paper is to expand human knowledge (and to improve the author's reputation); so the lecturer seeks to maximize understanding in the audience, while the academic author seeks to maximize his future reputation

A possible addition to your list: maybe scientists don't know their texts are impenetrable. I know I often skip important but "obvious" steps without noticing when explaining stuff. Lectures get repeated and can plausibly evolve to be understandable, while texts stay set in stone.

There are many texts origining from lectures and they are generally no more penetrable than texts written from scratch.

I am a bit skeptical about evolution of understandability. It seems to me that lecturers able to take a feedback and improve the structure of their explanations are rather rare.

Also, the audience in a lecture can (at least sometimes) give immediate feedback if they don't understand something, compelling the lecturer to explain in more detail.

Good point. It's certainly easier to quickly get feedback during a lecture. If academic writers really wanted to communicate understanding as much as [great] lecturers do, for this and other reasons it would certainly be more difficult to do so than it would during repeated lectures. I'm just skeptical that the desire is actually there to anywhere near the same degree though.

And it's not just a matter of different media. Consider a brilliant young researcher giving a seminar (i.e., spoken medium) on her research. Does she optimize for understanding or for making the strongest impression and convincing her peers that her research is important and original?

I agree and it depends on the lecturer, of course. Experienced lecturers seem to be more auditory oriented than aspiring ones.

There's also the pace of the learning to be considered: one sets his/her own pace when learning by text as opposed to speed and tempo chosen and changed by lecturer.

I think that a combination of both is the most effective way.

True, but if you ask any scientist about the primary purpose of papers, the reply will more likely mention communicating ideas (i.e. teaching - whether the information being taught is already known or is expanding human knowledge is hardly relevant for the style) than building author's reputation. So it's a kind of hypocrisy: authors say (think?) that his/her paper is written in order to inform, while in fact it is made to signal status.

It's not an attempt to 'signal status' but rather an attempt to signal brilliance and mastery of the subject, creativity, etc., and thereby tangentially to improve status (and reap the rewards of status, which might include non-status perks such as being able to research almost anything you want with large budgets, collaborate with the greatest minds in your field, etc).

I should add that I don't think this sort of dishonesty is conscious or intentional -- at least not in most cases.

Yes, you're right, I was simplifying too much (status and brilliance are strongly correlated, so I didn't pay attention to the distinction). However, this doesn't change much - still it's a hypocrisy.

Why do you think that the dishonesty isn't intentional? The journal reviewers demand a specific style and people generally know it. For most papers, the style matters if you want to get it published. If your status is enough high you are more able to choose your style, but a lower-status researcher should first maximise the number of his publications, and the most effective way is to imitate the approved style (which signals brilliance). The authors know it, they are not stupid.

The core of the problem is, in my opinion, that people really buy this signalling. If I don't understand what someone else is saying, I think (more or less automatically) that he's too smart rather that he's a poor teacher. Conversely, if I understand everything, I tend to think that the subject is trivial and the speaker is not much better than me.

I actually think it's good if a book asks me questions and lets me do some of the discovery for myself--conditional on the book having the answers somewhere, in case I find the puzzle too difficult or come up with an incorrect answer without knowing it.

I suspect that the informal nature of a talk might mean that one can describe things at an intermediate level of detail ("and so do this and that") that would look strange in a more formal paper, and so writers feel that they're stuck with either going into a full level of detail (which would be too much work to do for everything) or else saying "I'll leave this as an exercise to the reader"... even though going into some detail would be more useful to the reader.

I agree so much, I have to do it in writing.

Target audiences are different. More competent people prefer absorbing more dense presentation that leaves out obvious steps and communicates more sophisticated ideas faster, while newcomers need all steps spelled out for a simple enough version of the ideas. Video lectures are usually targeted as people new to the field, or generally less technically sophisticated, such as average university students. By looking for video lectures, you apply a heavy selection bias for this kind of target audience, and so you should expect the presentation to be more transparent. Written material, on the other hand, may target different groups of people. By searching for a random paper or book on the subject, you are more likely to encounter more difficult material. On the other hand, if the topic is popular enough, you should be able to find easy introductory written material as well, you'd just need to filter the material by less obvious cues than it being a video lecture course. One heuristic is to look for lecture notes.

That's not always the case. Plenty of times competent people are called upon to implement a new method, and want to see for themselves the precise steps that the techniques' discoverer has gone through. I don't always have time, and it's not always instructive to have to fill in the blanks.

I don't see any reason why more competent people would prefer omission of a key step in a proof with remark "left to readers as an excercise". If you want to save time, you can surely skip it while reading. Why shall the author make the choice for the reader?

Especially in math, even textbooks which are intended to present complete overview of some subject have these excercises to the reader. I like to read books from the beginning to the end, but this is exactly what makes it impossible for me to do with math books. It can be refreshing to stop reading at some moment and have to think about some detail, but to do it repeatedly, knowing that I am losing time trying to figure out something which could be explained on two lines, is frustrating. I read books and papers because I want to get the information as efficiently as possible, not because I want to eventually rediscover it for myself. Not much to do with competence, I believe.

Added: I also find "excercise to the reader" mildly offensive, as the author is saying that it is trivial for him, but not for the readers, who should improve their skills by making the excercise. Such sentences aren't surprising in undergraduate course textbooks, but when the audience are competent people, it seems inappropriate.

I find that going to the original paper generally does the trick. When an idea is new, the author will spell the details more carefully.

SilasBarta mentions difficulty with Boltzmann machines, Ackley et al.'s article is actually quite detailed, including a proof of the learning algorithm http://tinyurl.com/q4azfl

This is a maddening problem that's evident on even the lowest of levels. A while back I helped tutor a high school kid who was taking a geometry class online. His instructor had written electronic articles to supplement a textbook, but still did the same "skip a bunch of steps" nonsense that the textbook did, even though he had no real space limitation.

He (his mother, actually) asked me for help because (1) he wasn't a strong math student, and (2) he was absolutely stuck trying to follow some examples that purported to explain the first homework assignment.

When I tutored him, part of what we did was go through a couple of the examples micro-step by micro-step, identifying the assumptions and algebraic simplifications that happened along the way. This made him feel a whole lot better - he could grasp the new material more easily when it didn't take him 10 minutes of figuring out what old material had been already applied to the examples.

I write tutorials online occasionally, and try my best to not make this same mistake.

I mentioned in another comment that I have more commonly heard complaints of bloated textbooks, which drew the response from anonym that those aren't the books you're talking about. Can you give some examples of the books you're objecting to?

I would be most satisfied with an economic explanation. If lectures generally explain things better, then either a) the cost to lecturers of improving their explanations is lower than for writers, or b) the benefit to readers of a better explanation is less than the benefit to lecture attendees.

I would guess its the latter, because if an author loses you, you can google it. If a lecturer loses you, you might be lost for the next hour.

I think anonym gave us the Hansonian explanation: Texts are written to boost author reputations by seeming Formal, Authoritative, and Difficult; lectures are spoken to explain things. Though Hanson might or might not agree - see the discussion Hanson and I had about how much formality is signaling.

I agree with you and would just like to scream AAAARRGGGHHH!

There are enough factors at play that one can pretty much produce any conclusion one finds pleasing, by a suitable selection of observations and speculations. Here are a few more puzzle pieces for people to arrange into whatever theory they want.

  1. Many textbooks start as lecture notes, and for many others, their pre-publication drafts are used as the basis of lecture courses. Read their prefaces to find examples.

  2. A course of lectures is addressed to a narrow audience: one roomful of people studying that course, who are known to have the prerequisites for where the course starts. A book is addressed to a wide audience: anyone interested.

  3. You can have a conversation with a lecturer more easily than with a book.

  4. There are good and bad lectures, good and bad books, and good and bad web pages.

  5. In the civilised world, status is generally obtained by doing things that people value.

ETA:

6. Speech passes; text remains to be studied for as long as it takes.

7. Some people can explain things more easily in speech than in writing; for others it is the opposite.

Two corollary explanations come to mind. First, writing uses a wider variety of registers and styles than spoken language. Forms and usages that would sound exaggerated or affected in spoken language are socially appropriate in writing. Writing is constructed over time and predominately "for the record," so it uses precise, unforgiving language that suits the specific context of the writing. This is why the first line of a Wikipedia article on some topic in math, poetry, or physics is often indecipherable to a lay reader, even an educated one, without further reading. Spoken language, on the other hand, is first and foremost a form of communication from a speaker to a listener, and is composed and interpreted in real time, even if it's guided by notes. This makes it more fluid and colloquial, and more likely to employ a register that the speaker and listener will both understand readily. Since successful writers use the more precise, ossified language and successful speakers use the more fluid one, they diverge through memetic evolution, as suggested.

The second explanation has more to do with the way writing is taught. I don't know how much it applies to technical writing, maybe somebody can share their experience on that point. Since the Victorian era, prose has embraced brevity. The briefest explanation that still conveys the broad meaning of an author's idea is usually treated as the best stylistically. This sacrifices precision for a kind of clarity, but in a field like mathematics, precision is clarity. Typical admonitions about brevity of style, then, render useless attempts to explain big, scary concepts. Lecturers, however, have the opportunity to pursue digressions and explain minutiae in half-organized ways and still hold the attention of an audience because the lecturer can easily signal the importance of a difficult intermediate step to the wider narrative in a way that would be clunky and perhaps abrupt in writing.

It's especially bad in math

What is? the gap between lectures and writing? or just the amount of stuff left out?

My experience with math is that more steps are left out in lectures but the lectures are clearer. I agree with CatDancer that the difference is that informal high-level descriptions are allowed in lectures. That isn't really consistent with your specific story, though, since "do you see why?" is pretty informal.

So I suspect that you're mistaken. You think you want more detail, but the texts actually have more detail and it doesn't help.

On another note, along the lines of Michael Bishop and the commenters on anonym, lectures get more feedback than writings. I don't think lectures get a lot of feedback, but people do get a lot of feedback in a one-on-one informal oral explanations; and they give a lot more one-on-one explanations than they give lectures. That probably spills over into all oral presentations, making them more informal and better (perhaps independently). Simply by switching media, they lose a lot of that experience.

I'm quite sure that many people in academia just find it much easier and less time-consuming to write quick'n'dirty cue notes to themselves, then speak about the given subject for two hours at the time, than to render the same information in written form in easy-to-follow, clean language.

(Not to mention the fact that you'd probably have to do both; you're expected to lecture, after all.)

It often seems that when you see a paper presentation followed by a discussant, the discussant's two minute summary of the paper is clearer than the author's twenty minute presentation. Somehow it just isn't OK for the author to spell it out as simply as the discussant is allowed to. And as the post points out, written work is worse.

I don't think it's about what's OK to do (though see CatDancer), but that the discussant is more like the audience than the presenter is. The originator had one way of getting to the idea and this has colored his sense of what the hard parts are. The discussant and the audience are trying to learn it quickly, knowing what conclusions the originator reached.

[-]clay10

I always try searching for course notes that good professors put online. Did you try looking through Andrew Ng's SVM notes?

FYI to anyone who may be interested, the parent link is to lectures notes from Andrew Ng's awesome course on Machine Learning, available on youtube as well as through iTunes.

My experience is the opposite of yours in just about every point, and I have frequently seen complaints that science textbooks for American undergraduates are absurdly bloated.

"Bloat" seems to suggest that the author or publisher decided to put two or more books within the same binding...

There are different kinds of science textbooks, and the criticism of bloat is not leveled against the sorts of texts that are criticized for being too dense and impenetrable. For example, a common target of the bloat allegation is Stewart's 1368 page Calculus text. You won't find people making the complaints that Phil is making in this post about that book and others of its bloated ilk.

I ran into the same problem recently. A slashdot discussion referred me to Geoffrey Hinton's Restricted Boltzmann Machines, a method of unsupervised learning that he claims to have successfully applied to character recognition.

Despite reading several papers, a ppt presentation billed as a "gentle introduction" to the topic, and the video presentation, I could never discern what the actual algorithm is that Hinton is using. Fortunately, he posted the code and so I could look at the core algorithm (rbm.m).

But even then I couldn't figure out why exactly it's able to find structure in the data by working through toy problems (which it seemed to choke on). Eventually I gave up and started reading about Kohonen maps, which appear to do the same thing, but are much easier to understand.

OTOH, I could just be really stupid and unqualified to read this stuff to begin with.

Don't worry, you're not the only one who has difficulty reading Hinton's papers. They are quite confusing. The problem is that while there are only two or three big important ideas, they're spread out over the course of something like 20 papers.

This post makes me think we should have a LW math/ai/machine learning reading group. Choose a paper, read it, someone summarizes it in a main post, and then every one else chimes in with their 2c.

This post makes me think we should have a LW math/ai/machine learning reading group. Choose a paper, read it, someone summarizes it in a main post, and then every one else chimes in with their 2c.

Great idea! Also, make a list the links to these articles accessible from the Wiki. It would also be a great opportunity to apply the LW filters (as applicable to the situation) like:

-What regularity of the search space does the algorithm exploit? -Why does the algorithm appear to make use of randomness, and how can it be derandomized? -What phenomena does this theory rule out?

I'm sure there are some ways I could contribute, esp. with the Perceptual Control Theory stuff going around. I'm almost done with the four BYTE magazine articles on it that pjeby keeps referring to.

That sounds good, but in my experience with paper reading groups, most people don't find the time to actually read the material under discussion. It ends up being a venue for signalling and socializing. Maybe that wouldn't apply here.

I mean that authors are biased by the media they use.

McLuhan, of course. You've just expanded on the phrase "The medium is the message." Note that it applies far more widely than just things to write in.

[-][anonymous]00

I actually think there might also be a reader/listener bias. I sometimes feel that I've learned a lot more after hearing a lecture and less after reading a textbook, but when it actually comes time for me to manipulate the knowledge I find that the "incomplete" textbook presentation is often a lot more useful than the spoken lecture which I find to be full of gaps that require me to think through things. This may be because my work often requires ideas that I generally only find in rigorous details of the algorithm or proof and such details generally would lose and audience so they are omitted.

Also, to say that in a five-hour lecture series one could cover the "mathematical details" among other aspects of the SVM means that you'll miss out on a lot of important topics! I use them a lot, have read over a thousand pages on them, and have gone to countless lectures, and I'm still learning what they are about.

What are we to do to remedy this?

What are we to do to remedy this?

Write the books we want to read.

thanks for the link to videolectures.net

I think that there is also the perceived friendliness that goes both ways.

+The author seeing audience helps in giving feedback. It is easy to leave out hard steps when you do not have to see the blank faces.

+The audience also benefits by increased communication through non-verbals signals. Such as cues in what parts are extremely important and the like.

I think the real reason is lazyness. From my own experience writing short articles on the internet(like this one) I know that I often will leave out the details and write only the essentials for brevity's sake. It's much easier to talk than to write and so while talking you can go into much more detail and you also have all the subcommunication channels(gestures, facial expressions, variation in tonality) at your disposal. Plus while writing you have to take care of forming correct sentences, etc... so it's more effortful.

To contain all the details a text would probably be much longer.

Edit: Also remember that we have native hardware that enables us to speak, writing on the other hand has to be learned and doesn't come off as easily.

Edit2: Even while writing this short comment I read it at least two times to make sure it's not too bad, yet I still feel embarrassed because I'm not a good writer, whereas if I could speak the same comment I would long have finished and moved on to something else.

Edit3: Adding this little notes to my comment is also more effort, I have to bring up the page again, find my comment, press the edit button, write, make sure that it is intelligible, and then press the comment button again. Again, talking is much easier. I hope I got the point across. :)