Wiki Home

Linear probing

Definition

A linear probe is a linear classifier trained on a network’s intermediate activations to predict a property of the input, used to test whether that property is linearly decodable at a given layer (Alain and Bengio, 2017). High probe accuracy shows the information is present and linearly accessible; on its own it does not show that the network uses that direction in its computation, a standing caveat in the probing literature (Belinkov, 2022). The weight vector of a trained probe is itself a concept direction, which connects probing to methods that read or edit concepts along directions in activation space.

Papers that use this concept

No reading-companion page currently uses this concept.

See also

• Linear representation hypothesis
• Activation steering
• Mechanistic interpretability

References

• Alain, G. and Bengio, Y. “Understanding Intermediate Layers Using Linear Classifier Probes.” International Conference on Learning Representations (ICLR), 2017.
• Belinkov, Y. “Probing Classifiers: Promises, Shortcomings, and Advances.” Computational Linguistics, 48(1), 2022.