OptimNN: neural-network parameter generation (hypernetwork) for optimizing BKT¶

Theory · All theories
Evidence · no claims cited

Description¶

OptimNN replaces random initialization and direct SGD training of BKT's four per-skill parameters with a feedforward "parameter generation network" f that maps a skill ID to the four BKT parameters, so gradients update the network rather than the parameters themselves. The paper describes it as "similar to hypernetworks (Ha et al., 2017) or black-box adaptation meta learning", an equivalent re-expression of the BKT optimization problem. It reports benefits including improved numerical stability, vastly reduced hyperparameter tuning, flexibility in preventing parameter degeneracy, and empirical performance improvements over SGD and existing optimization methods.

Design Implications¶

Context¶

Requirements¶

• A deep learning library with automatic differentiation to compute gradients with respect to the generation network's parameters

Constraints¶

• The generated parameters are limited to the optima attainable within the set of all BKT models, which motivated the BKTransformer extension

Target Learners¶

• students in intelligent tutoring systems and computerized tutoring contexts

Target Learning Objectives¶

• estimating student skill mastery and predicting response correctness in knowledge tracing

Claims¶

• Optimnn Reg Minimizes Degenerate Parameters [+M]

Related Theories¶

• Skill discovery BKT: end-to-end stochastic learning of the problem-KC assignment matrix with Gumbel-Softmax
• Bayesian Knowledge Tracing (Two-State Hidden Markov Model)
• Bayesian Knowledge Tracing: a four-parameter student-learning model in two forms (HMM and Knowledge Tracing Algorithm)

Examples¶

-

Key Sources¶

• Badrinath, A. and Pardos, Z. (2023). Optimizing Bayesian Knowledge Tracing with Neural Network Parameter Generation. https://github.com/abadrinath947/OptimNN