4.3 Metalearning and autoML

.... special slides ..... yikes

Variety of Models

More specific knowledge, that is, models for smaller subsets, eventually, individual entities

Model Management

Bias

Can't live with it: given a dataset and a learning algorithm, not every model is possible
Can't live without it: bias-free learning is futile
Bias-free algorithm: can learn any model but doesn't have any preference for one over the other

Metalearning/automl Role

Metalearning for algorithm selection
- induce model from metadata to predict the best algorithm on a new dataset

Opportunities and Challenges

Dataset morphing to understand ML algorithm behavior
dataset characterization
dataset embeddings

< Go back