Deep Learning is Not Always the Best Solution in Education

Building AI models in education seems like a slam dunk. Educators and Edtech companies are so eager to make use of this new technology that sometimes the resulting product misses the mark. Deep learning has a place in education, but to make the best use of these types of models, educators and innovators need to slow down and consider what its best use is—to realize it’s not always the best solution in education. 

Dr. Josine Verhagen from Kidadaptive has some tips for knowing when deep learning is appropriate and when educators should look to other models in her talk for ODSC West. Let’s take a look at her advice for nailing deep learning in Education.

[Related Article: Deep Learning Finds Fake News with 97% Accuracy]

Where Deep Learning Shines in Education

Deep learning is an exciting development in education. However, the problem you’re solving must be suitable for deep learning. Here are some critical markers for deep learning:

• a large dataset that’s actually representative of the target data.
• sophisticated features that must interact with each other (speech recognition for example)
• clear outcome to predict that’s relevant for your learning goals.
• the outcome is more important than model transparency.
• the need to optimize data for specific situations that won’t transfer
• enough time to do the training

For example, in language learning software, deep learning is appropriate for pronunciation training. Verhagen’s team created a program to help train children in a natural environment where they can speak, and the software recognizes if their speech is recognizable in the target language.

Data Availability

To train a deep learning model, you need enough representative data to inform accuracy and predictability. Your features need to have complex interactions so that your program learns from the child or learner through multiple types of actions. If you aren’t providing this type of rich environment, going through the trouble of deep learning may not be worth it.

In many of these programs, the only recorded information is a correct or incorrect answer. You might have a record of how much time it took. The relationship between the features is straightforward and measuring when the student has actually learned something is difficult. The limited sequences of responses don’t lend themselves to deep learning necessarily because the features aren’t sophisticated enough.

We’re interested in more than the predictive modeling of “will the student get the next question correct?” The models don’t account for human behavior such as the chaos of working with preschoolers, and our outcomes won’t actually predict what the learner will do. The features don’t align with the checklist above, and usually, that means deep learning isn’t going to provide the results you want.

Understanding Your Model

Educational tools often need transparency. If part of your assessment requires an explanation of how that decision was reached, deep learning won’t be the best option. When predictions have real-life consequences, it’s best to look for other solutions than deep learning. That transparency also extends to the learner. If the learner needs to understand and apply those results for actionable outcomes, deep learning won’t be constructive.

If you need to reuse the model for similar actions, deep learning could be prohibitively difficult. The training involved in pivoting to another situation would take up a lot of time and prevent you from deploying your educational model as quickly as you want.

Deploying Deep Learning

Deep learning has exciting applications in the world of education, but it’s not a cure-all for every educational problem. In areas where transparency is paramount, it could be challenging to make a case for the hidden logic of deep learning. Your students will need more advice and actionable results that deep learning can’t offer.

[Related Article: How to Leverage Pre-Trained Layers in Image Classification]

Instead, for models where the prediction itself is the outcome, and your goal is to use the results as a tool rather than an actionable result by itself, deep learning can classify that complex data and make it available to people like educators or administrators who need that sort of data. Otherwise, sticking to simpler models with basic machine learning algorithms will save you the time and headache of implementing a model that might ultimately prove ineffective at best and unethical at worst.