Saturday, October 27, 2018

#el30 Data and Models

I should be grading student documents this morning, but I'm thinking about #el30. I may have an assessment of that next week.

Anyway, as I was reading some posts about Data, I was struggling with our previous discussion about the differences between human and machine learning, when something that AK wrote sparked some coherent ideas (at least dimly coherent for my part). AK said: "This got me thinking about the onus (read: hassle) of tracking down your learning experiences as a learner. ... As a learner I don't really care about tracking my own learning experiences."

I thought, no, I, too, don't want to track all my learning experiences. Tracking all those experiences would take all my time, leaving no time for more learning, much less time for grading my students' papers. So maybe computers can be useful for tracking my learning experiences for me? A computer can attend me--say, strapped to my wrist, in my pocket, or embedded in my brain--and collect data about whatever my learning experiences are. After all, computers can collect, aggregate, and process data much faster than I can, and as Jenny notes, computers don't get tired.

But what data does a computer identify and collect? Even the fastest computer cannot collect all the bits of data involved in even the simplest learning task. How will the computer know when I'm learning this and not that? Well, the computer will collect the data that some human told it to collect. Can the computer choose to collect different data if the situation changes, as it certainly will? Perhaps. But again, it can only ever collect a subset of data. How will it know which is the relevant, useful subset? The computer's subset of data may be quantitatively larger than my subset, but will it be qualitatively better? How might I answer that question?

Turning experience into data is a big issue, and I want to know how the xAPI manages it. Making data of experience requires a model of experience, and a model always leaves out most of the experience. The hope, of course, is that the model captures enough of the experience to be useful, but then that utility is always tempered by the larger situation within which the learning and tracking take place. Can a computer generate a better model than I can? Not yet, I don't think.

If both the computer and I are peering into an infinity of experience, and I can capture only about six feet in data while the computer can capture sixty feet, or even six hundred feet, we are both still damned near blind quantitatively speaking. Reality goes a long way out, and there is still something about constructing models to capture that reality that humans have to do.

I've no doubt that computers will help us see farther and wider than we do now, just as telescopes and microscopes helped us. I've also no doubt that computers will help us analyze and find patterns in that additional data, but I'm not yet convinced that computers will create better models of reality without us. When I see two computers arriving at different views of Donald Trump and arguing about their respective views, then I might change my mind.

No comments:

Post a Comment