Victory’s spinoff metacog has been busy adding new features and functionalities. When companies look to incorporate metacog into their digital products, they want to know two things:
- How does metacog work?
- What can metacog help me do now that I couldn’t do before?
The answers to both questions lie in our unique approach to guided deep learning: machine learning steered by an understanding of real student outcomes.
In education, deep learning is different from deeper learning, which is a pedagogical approach to instruction. In the world of Big Data, deep learning is an artificial intelligence (AI) approach that creates neural networks in which each “neuron” is a computer processor. This structure mimics how the human brain works in parallel processing.
Deep learning can be very effective, but it has a drawback: neural networks are so complex that we can’t know how they arrive at certain decisions.
Guided Deep Learning
At metacog, our guided deep learning process begins with a clear definition of what constitutes a good result. The computer program then goes on to do the heavy lifting. We can trust the reasoning behind the program’s decisions because we supply the reasons!
Without proper guidance, deep learning on its own can be problematic. Take self-driving cars, for example. They can use neural networks to observe and mimic human drivers, but unless the software can also effectively evaluate proper behavior, the resulting decisions can be erratic, and even dangerous. We don’t want self-driving cars that emulate a driver who is distracted by a text message!
metacog’s guided deep learning approach, on the other hand, specifies how to measure the appropriate outcomes. Like a self-driving car that models only the best driving behavior, our system understands the goals users want to achieve, and knows when those goals have been met successfully. How is that achieved?
Rubrics to the Rescue
The measurable goals are defined by rubrics. Just as rubrics are used to guide a teacher in scoring an open-ended performance task, rubrics are used to guide metacog. We set up a simple system in which humans create a rubric that defines what behavior constitutes a good score and what constitutes a bad one.
Once the rubric is defined, the deep learning program is then guided in how to apply the rubric. This is done with training sessions. In one training session, an educator scores the performance task while watching a playback of a student’s performance. Given enough training sessions, the deep learning program can then emulate the scoring with good accuracy and impeccable consistency—just like a self-driving car would emulate a perfect driver.
The benefits of machine scoring are enormous. A student can get immediate feedback while doing the performance task, instead of waiting for the teacher to grade the performance task after they have finished. And a teacher can be alerted immediately when a student is struggling, and then take appropriate steps for remediation. So metacog does not replace the teacher. Instead, it puts a highly intelligent teacher’s aide at the side of every student.
If this topic piqued your interest, here are a few links for a deeper dive.