Scalable Nonlinear Learning with Adaptive Polynomial Expansions
This work addresses the challenge of scalable nonlinear learning for machine learning practitioners, though it appears incremental as it builds on existing linear representations.
The paper tackles the problem of learning nonlinear representations with computational efficiency comparable to linear methods, resulting in an algorithm that shows favorable computation/prediction tradeoffs against strong baselines in experiments.
Can we effectively learn a nonlinear representation in time comparable to linear learning? We describe a new algorithm that explicitly and adaptively expands higher-order interaction features over base linear representations. The algorithm is designed for extreme computational efficiency, and an extensive experimental study shows that its computation/prediction tradeoff ability compares very favorably against strong baselines.