Deep Learning 2.0

The net is often criticized for enabling shallow learning, while leaving deep learning largely untouched. It’s easier than ever to pick up simple facts and trivia, recipe style techniques, little snippets of information. But, if you want to learn the equivalent of an undergraduate chemistry course, it’s just as much slog as it ever was.

There are people who do deep learning differently though. Autodidacts have the skill of being able to sift through structured texts and courses in what should be an area that requires long, slow deep learning, and “pick out the eyes”, grabbing just the knowledge they need. The problem here is that it requires some really unusual skills and abilities to do that, because the materials actively get in the way of this approach.

There is no reason we can’t make this easier using the ‘net. There are masses of deep learning content available for free online, it’s just impenetrable to a greater or lesser extent. If it was structured appropriately for self learners, we might get a modest renaissance going!

I propose a simple structure for representing the essence of deep, structured learning materials, in a way that makes them usable on an ad-hoc basis.

There are three fundamental entities in the structure.

– Know
– Learn
– Person

A “Know” is a description of an atom of knowledge/skill/understanding. It could describe anything. EG:

Know Name: WordPress-Blog-Posting
Know Description: Understanding of and ability to post to a WordPress Blog

A “Learn” is an atom of learning. It describes how to acquire one or more Knows.

Learn Name: Tutorial on How to post to WordPress
Learn Description: Refer to the tutorial at . Run through the whole thing, and there are some self test questions at the end if you want to make sure you understand.
Approx Duration: 20 min
Depends on these Knows: WordPress-Blog-Signup, Basic-Browser-Use
Delivers these Knows: WordPress-Blog-Posting

A “Person” is a description of a person in the context of the system. In particular, the system tracks a list of Knows that the person says they have, and possibly a list of Learns that the person has undertaken.

Given this simple three element system, all kinds of use cases are possible.

– Acquire “Know X”
The Person wants to acquire Know X. The system traverses the graph of dependencies, between Know X and the list of Knows the person has already acquired. The system can then recommend a path from where they are to acquisition of Know X, as an ordered list of Learns, just like Google Maps can give you directions from point A to point B. There can be a set of alternate recommendations.

Then the user just visits each Learn when ready, follows the instructions, and works until they feel they have acquired the Knows.

– Add to the system
People can simply add Learns and Knows as they see fit. Know and Learn names are tag like; you can define them as you wish, and people can attach their learns to your knows if they like.
Maintaining existing knows and learns could be wikipedia-like – anyone can edit, but there’s a system of admins to sort out problems. Or, they could all be owned by their creators, and if you don’t like what a person has created, you can create rival Knows and Learns that you think better reflect reality.
The dependency system (Learn X requires Know Y and gives you Know Z) might also be vote/like based; anyone can propose new dependencies, and give a thumbs up/thumbs down on existing ones, so over time you get a good relative sense of which dependencies are truly important and which aren’t.

– Prove your knowledge
A social relationship between Persons, or just public profiles where desired, could let others see the “Knows” you have.
As an extension, a Know could be owned by someone who reserves the right to allocate the Know to other people. This means they take on the responsibility of verifying that any given person knows their Know, which might be testing based, or based on records of use of a third party system, or even based on undertaking a university or other private, commercial course. These kinds of third party verified Knows could become resume items in the real world. Of course, this could hamper autodidacticism, but then anyone can create alternate, unverified routes around such Knows, and/or set up an open, free verification system. A simple peer review approach could also be implemented into this system itself as another extension. Note also that this is all secondary; access to structured learning tools themselves is the primary thing.

This all needs more thought. Firstly, it needs to be a self-managing, “crowdsourced” system. Success would rely I think on getting a system of reputation exactly right, so that reputation based incentives drive engagement.

Knows and Learns need to be the smallest possible atoms. If you see a Know or a Learn and think “That’s too complex, I could break that into parts and associate them with dependencies” then that should be done. Approx Duration should ideally be on the order of a few days at the most, hours is better.

There probably also needs to be a “combine” mechanism, maybe more like “subsume”: take Know X, and subsume it into Know Y (change all dependencies on X to be on Y, maybe add in some of the info from X to Y, then delete X).

Some aggregating mechanism might be required, but then simply making aggregate Knows might suffice. If we just add “Equivalent To” to Knows, and allow a list of Knows (or a list of lists of Knows, so we can have alternatives), then the Know “Chemistry-Major” could be created, with a short description, and an “Equivalent To” list specifying all the component Knows. This could be multiple levels deep.

If this is successful, we should see people going out and “picking the eyes” out of the free online materials: MIT OpenCourseware, sites like Academic Earth, free online textbooks, how-to sites, youtube instructional videos, informative blogs, you name it, creating interlinked Knows and Learns.

The first version should be stupidly simple. A page for Knows, a page for Learns, some basic editing facilities, some basic searching/nav facilities. Tracking what Knows you’ve acquired probably isn’t necessary first up. Editing facilities may be more important than use early on, and should assume an intimate group that communicates and self polices, suitable for a very small group of editors. As it grows, the system should change to match changing dynamics, but up front, simple unrestricted editing might be best.

That’s all I can think of for now. Comments?

Deep Learning 2.0

2 thoughts on “Deep Learning 2.0

Leave a Reply

Please log in using one of these methods to post your comment: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s