Learning Abstract Task Representations
This work addresses the challenge of data characterization for algorithm selection and performance estimation in meta-learning, offering an incremental improvement over existing meta-feature methods.
The paper tackled the problem of improving meta-learning by learning abstract task representations as latent variables in a deep neural network, resulting in induced meta-models that outperform other methods by ~18% on average in mapping to generalization performance.
A proper form of data characterization can guide the process of learning-algorithm selection and model-performance estimation. The field of meta-learning has provided a rich body of work describing effective forms of data characterization using different families of meta-features (statistical, model-based, information-theoretic, topological, etc.). In this paper, we start with the abundant set of existing meta-features and propose a method to induce new abstract meta-features as latent variables in a deep neural network. We discuss the pitfalls of using traditional meta-features directly and argue for the importance of learning high-level task properties. We demonstrate our methodology using a deep neural network as a feature extractor. We demonstrate that 1) induced meta-models mapping abstract meta-features to generalization performance outperform other methods by ~18% on average, and 2) abstract meta-features attain high feature-relevance scores.