Deep learning is not always the best solution: Illustrative examples from educational products

Abstract: Deep learning is rapidly taking off as an incredibly useful tool for image classification, NLP, and many other applications. However, not every data science problem is served by a deep learning solution. This talk will explore several disadvantages of deep learning models and propose alternatives based on three applications of data science to data from educational products.

First, deep learning models often lack transparency about how predictions are generated. That transparency, however, is essential to the success of an educational product which recommends, gives insights and makes predictions for students. When providing learning reports to parents, teachers, or students themselves, results from non-transparent models can decrease trust in the product. Second, prediction accuracy is not always the goal. It can be more interesting to look at what happens when the prediction is wrong. Having knowledge about the predictive model beyond accuracy of prediction will help to provide more insight into the learning process. Specific theory driven models (e.g. from psychometrics) often give equal or better results than generic machine learning models by modeling known relationships between features. Third, in the absence of thousands of features, deep learning models often end up being a time and resource expensive way of estimating a simpler model. With a limited set of features, simpler models are often equally accurate in prediction, less time intensive, and easier to interpret.

At the end of this talk, the audience will have a better sense of alternatives to deep learning models in cases where model transparency is important, where prediction accuracy is not the main outcome, and where the feature set is small.

Bio: Coming Soon

Open Data Science Conference