[snikolov]

Just-in-time learning is the future. No ifs, ands, or buts. Any argument you might have to the contrary is not only wrong, but dangerously wrong ... The current system of teaching students everything they might need to know, just in case it ever comes up, is an artificial construct that resulted from limited access to books and a limited amount of time available for study. Just in time, on the other hand, is how we actually learn by default.

This is an interesting point. I happen to disagree, and I hoped that the author would go on to say exactly why it's wrong, especially dangerously so. It seems like the author rejects curriculum-based courses of study that provide a broad, solid foundation in favor of just-in-time courses of study, in which one learns whatever is necessary at the moment, right before applying it.

My main objections are that

* There are things you don't feel like learning that you would do well to learn. I had a lot of freedom to choose courses during school, which was great, but I am about to graduate with a lot of holes in my knowledge.

* No matter how smart you are, you would benefit from the guidance of a teacher --- guidance through a full curriculum that gives you a solid foundation and imparts onto you important patterns of thought.

[vacri]

This was my sticking point as well. University doesn't just teach you facts, it teaches you how to think constructively, liaise, and how to research. Who is going to say "I need JIT Learning to teach me the scientific method"? That's an example of something that most people need to have introduced to them; folks just aren't aware that it exists as a process, let alone why.

[padwiki]

I was hoping this comment would come up, actually. The real difference in just in time vs. just in case has to do with delivery, sequencing and retention. A broad, solid foundation sounds great on paper, but there are some hard realities that sink in shortly after graduating. First, within about 6 months the amount of actual knowledge you retain from this foundation is probably going to be between 5 and 7%. That's 5-7% of the total from a not very efficient 4 or 5 years of dedicated study. How not efficient? Well, if you look at the in major portion of your degree, you probably had about 450 hours in real classroom instruction, or a little under 2 hours a week. So, of those 450 hours, you will probably retain less than 30 or 40 hours worth of real useful knowledge. you may dismiss this as hyperbole, but trust me, all that work dissolves incredibly quickly. I remember Big O, a couple of formulas, and the ability to order a sandwich (or was it a donut) in German.

More importantly, though, is the fact that since you only have 40 or so classroom hours in any class, and you have to teach to the average, it is extremely difficult to build up to a properly high level of skill in any particular subject. It feels hard while you are doing it, but after you graduate, you realize the people who have been focusing on the subject for a couple years are light years ahead of you. It's even more difficult to chain subject together to reach that high level. The closest thing we have is a generic 100,200,300 level system with some prerequisites.

How does this relate to just in time vs just in case? Even if you assume an identical breadth of knowledge, being able to sequence classes together in series instead of having semester and scheduling gaps means you go into the next class with more knowledge retained from the previous, which means you can build on your foundation in a more logical and efficient way and reach those higher levels that you just can't in a fragmented system. You can approach this from the ground up (building on higher and higher concepts), but the very nature of a JIT system means you can also approach it from the top down. That is, you can define the ends result or top level class, and then sequence each course to build up the fundamentals you need, just before you need them.

The point? If you are defining a broad base of skills, JIT allows you to master each one quicker and sequence them together to reach higher levels of mastery. If you need skills in the real world, JIT is the quickest and most efficient way to build those skills. The reason I consider disagreeing to be dangerously wrong, is that JIT is so much more effective at real education that those who bank on JIC for their future (students, schools, or countries) will find themselves left in the dustbin of history.

[snikolov]

Thanks for the thorough reply. It seems we agree on a few key points:

1) Retention sucks in the current model of higher education.

I've been thinking about this one a lot lately. I've been doing a one year masters where I'm taking two courses a semester and doing research. The depth with which I am learning things is night and day compared to the depth with which I learned my undergrad material. During undergrad, I was drinking from a firehose and just trying not to drown. I would turn in unfinished problem sets, not having learned the material, and move on with my life. I would sleep through classes out of sheer exhaustion.

Now, with just two classes, I'm able to learn things almost well enough to teach them. So one way to improve retention is just to take things slower.

Another model comes to mind if we consider how people study at Cambridge, Oxford, etc. I have not experienced it myself, but according to students who went there as exchange students (and students from there who came here (here being MIT)) it's pretty different. Students here are overwhelmed with constant work. There, it is a lot more self paced, with a set of final examination at the end (someone please correct me if I am not doing it justice). So perhaps self-pacing and working smarter, not harder leads to more retention.

Do you know of any sources for retention statistics such as those you cited? Some of them don't match my experience (for example, I would say that I spent 8-10 hours a week in classes related to my major).

2) Courses need to happen in a logical sequence so that they can build on one another.

When I first read your post I thought you were suggesting that students should, by themselves, pick what to learn based on what they want to build, in lieu of being guided through a logical curriculum.

The point about scheduling gaps is interesting. Scheduling gaps happen because it's hard to satisfy the constraints of so many student and faculty schedules. If you could take courses on demand, that would fix things.

3) Results driven learning can be excellent for motivation and retention.

When one talks about results, there is a fundamental issue of time scale.

Courses that say things like "When you're done, you will have built an autonomous mobile robot" are great.

But there are many fundamental things to learn, over a long period of time, whose benefits

* you might not see for a long time * are broader than you could have ever imagined (and hence the benefit would seem artificially low to you)

If you as a student get to pick the desired result yourself all the time, you might be tempted to pick shorter-term results. This can be catastrophic to your education.

I believe in forcing people to learn fundamentals of their chosen field --- fundamentals whose power they might not appreciate until later. Learning fundamentals (that you might choose not to learn if you weren't forced to) is fruitful in powerful and unexpected ways.

Take pure math classes. You learn analysis. Then you learn measure theory. Then you learn measure-theoretic probability theory. Then you learn stochastic processes. All of a sudden, financial mathematics becomes easy to grasp. But so do a host of other things. Signal processing, computer vision, statistical mechanics, complex multiagent systems, epidemic modeling, control systems.

I suppose you could have started on this path because you wanted to learn financial mathematics. But it probably would have seemed way too complicated and difficult. But if someone says they want to be an applied mathematician (a much "broader" and more long term goal than just learning financial mathematics), then they'd better take a ton of pure math.

[padwiki]

1) There have been a few different experiments with self pacing and retention at schools like Oxford (and a college in Iowa of all places). The typical result is a significant increase in retention with lower stress. The problem, and fundamental reason why this isn't the norm, is simple logistics. Trying to build a serial system of education that still uses classrooms is massively difficult. The closest you can get is a compromised 6 week class system that is still extremely difficult to pull off. Parallel is just easier and more cost effective to administer. It's also easier and more cost effective to administer on an online system which is why the push for MOOC. The difference for us is that we care more about optimizing the learning process itself and real evaluation and guidance, and are willing to bet that enough students also care about those points to give us a chance at proving it at a large scale. Our system costs a little more than free, sure, but the difference in results can be dramatic, as you are seeing with your masters program.

I don't have the 7% study at hand, but I'll try to look through my list of references to see if I can track it down. They cover some statistics on retention in this article: http://www.wired.com/medtech/health/magazine/16-05/ff_woznia...

As for in-major hours, it may vary a bit by school (and MIT has a higher percentage than most), but here is the breakdown for UT. In-major credits required for a CS bachelors: 45. Average credits per class: 4. Weeks in a semester: 14. Average classes per week: 3. Total estimated lecture hours per class: 42ish. Total in major lecture hours for 45 credits: about 500. Subtract intro days, quiz and final days plus time for class to get settled, etc.. and you have somewhere around 400-450 hours left, or 100 per year, 2 per calendar year week or 3.5 per school year week. Some semesters you might not have any in major, some you might have 2 or 3, so remembering 8 hours in a week of classes is not outside of the data set.

2) A student should not have to know the exact set of skills necessary to build up to the higher level concepts. This is one of the problems with the self directed learning attempts in traditional schools. But, if the student has a clear end goal in mind (say, the goal of becoming a search engineer at Google), then a JIT system that can build a sequence of courses based on the concepts within the course is more effective and targeted than a standardized fire-hose curriculum. Speaking of fire-hose, I'm working on a post covering context switching in a parallel educational system that addresses exactly why this is such a problem.

3) This comes back around to the top down (or goal oriented) approach, and gets even more interesting with a mentor guided approach. Imagine a hybrid apprentice/mentor system where the mentor can define a sequence (or collection) of high level concepts which the learning system could then take and generate a path to master those concepts. All the while, the mentor could provide direct guidance and assistance where needed and help answer the tough questions that arise without having to dedicate all their time to the actual instruction. There are many, many, variations on this pattern that work extremely well with JIT systems but really don't work at all in JIC.