How Deep is Deep Learning?
summary
Compares Deep Knowledge Tracing (DKT) — an LSTM-based deep learning model with hundreds of thousands of parameters — to classical educational models Bayesian Knowledge Tracing (BKT) and Performance Factor Analysis (PFA) on a large 5M-point dataset from the funtoot tutoring platform. Shows DKT outperforms BKT but matches PFA, analyzes why BKT lags and how enhanced BKT can reach DKT performance, and argues that the knowledge-tracing domain may be shallow so deep models offer little practical advantage despite higher complexity.
AI-written summary · about
chapters (8)
- 0:00Start
- 1:05Homogeneous Teaching
- 1:55Two Sigma Problem
- 5:30How do we approach this?
- 6:29Deep Knowledge Tracing (DKT)
- 9:46Inter-skill Relationships
- 11:39Results. Accuracy
- 16:33Deep Model: Advantages
related talks
31mHow to efficiently model learner’s knowledge with...
45mIntroduction To Deep Learning
40mHuman Learning
51mTensorFlow Wide & Deep: Advanced Classification the easy way
37mRecent Solutions Using Deep Learning Models/ The How and Why
01h 04mWhat you always wanted to know about Deep Learning, but were afraid to ask
48mML and DL in Production: Differences and Similarities
37mDeep Recommender Systems