Frozen Feature Probe
A standard paradigm in representation evaluation in which the weights of a pretrained model are frozen and only a lightweight linear classifier or regressor is trained on its output features, measuring the amount of linearly decodable information contained in the features.
Timeline (last 90 days)
The experiments used frozen-feature probes to evaluate ten representations under streaming egocentric observations in static indoor environments.
The limitation of frozen-feature probes is that they can only show that information is present and readable, and cannot prove that the model actually uses this information in downstream reasoning.
All Facts (2)
The experiments used frozen-feature probes to evaluate ten representations under streaming egocentric observations in static indoor environments.
50%UnverifiedThe limitation of frozen-feature probes is that they can only show that information is present and readable, and cannot prove that the model actually uses this information in downstream reasoning.
50%