For the past several weeks I have been working on an Independent Study course at Brandeis University which I have titled Exploring the Exocortex: Machine Learning for Human Behavior, advised by Professor Tim Hickey.
Originally conceived as an attempt to use biologically inspired machine learning techniques such as neural nets and genetic algorithms towards modeling and then improving day-to-day human behavior, the course has moved towards a more direct path to solving that problem.
I have read several papers and chapters in books, summaries of which I will post soon. In the end, this series of posts (which my be followed using the category exocortex on this blog) will adapted and augmented into a paper, which I will also post here. I believe that research should be done openly and publicly, and so, that’s just what I shall do.
What is the exocortex?
To the best of my knowledge, this is a term coined by researcher Ben Houston–and popularized by science fiction author Charlie Stross–to describe the various systems humans may use in thinking but which are not part of our bio-brain. Already, our Blackberries, iPhones, and other essential electronic devices are proto-exocortices (yup, the plural isn’t pretty).
Why am I working on the exocortex?
As human civilization has grown, we have increased in complexity. Some welcome this, some don’t. Some believe that it will lead to some sort of Singularity. The Flynn Effect most likely is a result of humans attempting to adapt to this environment which is growing exponentially more complex. Already the problems of an Attention Economy, pioneered by the same people who pioneered modeling human behavior and augmenting human cognition, are apparent: There are more things one must pay attention to, within the same time constraints and physical limitations.
Thus, it seems obvious to me that to cope with this information, and more importantly, attention load humans must create appropriate tools. The exocortex is a collective name for those tools.
What do you mean by “Human Behavior”?
I am planning to specifically tackle the problems I have greatest difficulty with allocating attention to: those pesky appointments and other thing one might put on a calendar. These things have a relatively high importance, and also allow pretty easy assessment of goal completion.
How do you plan to work on that?
I am devoting this course to creating a document detailing what I believe is a path of least resistance to a piece of software that can model strategies for goal completion and evaluate the best ones. If time permits, I will implement as much of it as I can.
Here’s a rough plan for such a system:
- Goals are inputted into the system.
- User provides goal strategy by narrating real-life activities.
- User strategy narrative is annotated using frames from FrameNet.
- Goal-completion satisfaction is rated by user. This is somehow applied to constituent frames.
- Process is repeated, and different frames are assigned different valuations based on perceived contribution to goal completion.
- System provides best set of frames to form optimal strategy for the completion of each goal.
How does this system differ from a PIM, e.g. on a Blackberry?
Various calendar systems may provide reminders, perhaps with some intelligence noticing your location etc., that assign to you the task of evaluating your current strategy and seeing if it matches a hypothetical optimal strategy for accomplishing the goal specified in the reminder. This is an attention heavy process. Instead, I would like to move as much of the strategy modeling and evaluation as possible out to the exocortex. Even when strategy evaluation is not yet optimal within the system, merely providing concrete strategy options should reduce the attention needed by the user to evaluate a course of action.
Are you really going to do this, and not let it stagnate like you’ve done with Gargoyle?
Well, I’m still working on Gargoyle, slowly but surely! Many new things in my life have taken time I could spend on it (some will be revealed soon).
This exocortex project, however, is guaranteed within the semester time frame as a grade depends on it. So, you can be assured of results. I hope you’re interested and excited, because I am!
Popularity: 88% [?]
3 Comments
I will believe that your system really can optimize strategies if it could resolve the following problem without having a specific algorithm for this kind of problems.
Here is the problem?s context.
You?re making French toasts and the recipe calls to fry each toast on a skillet exactly one minute on each side. Your skillet has room for two toasts only. The question is: ?What?s the minimal time needed to serve five toasts??
The system isn’t meant to infer new modus operandi. rather, it selects the best of what you’ve done. it essentially optimizes the task of “learning from your mistakes”. With a formalized way to define goals, you could also learn from other peoples’ mistakes. That is, if you can define “make french toast” formally, you can share strategies over the network.
So, your problem definition isn’t valid within the context of my proposal. Once a human solves the problem you described, the system would recognize when you are deviating from it, and let you know “oh, whoops, you’re not doing it the good way!”.
So it seems that at least three aspects are not clear for me.
First, if my solution differs from yours what solution your system will choose (especially if we both came to the same result by different ways)? BTW, what’s your solution to the presented example?
Second, how we will enter our respective approaches into the system?
And the last (but obviously not least), when one fries French toasts, how the system will correct him/her on the way?
Post a Comment