Representation learning
How does a model decide which features of a complex signal matter?
Example experiment
One way to probe a learned representation is to freeze the encoder and test which attributes a simple classifier can recover.
How do machines learn structure from data, and when can we trust what they learn?
Explore research ↓How does a model decide which features of a complex signal matter?
One way to probe a learned representation is to freeze the encoder and test which attributes a simple classifier can recover.
How do relationships between observations change a prediction?
Compare a graph model with a feature-only baseline, then shuffle the edges to test how much information the connectivity contributes.
What happens to a model when the data change?
Hold out a collection site or time period and compare performance and calibration with the original test set.
Asst. Prof. Cemre Candemir