Alert button

Constraining Representations Yields Models That Know What They Don't Know

Aug 30, 2022
Figure 1 for Constraining Representations Yields Models That Know What They Don't Know
Figure 2 for Constraining Representations Yields Models That Know What They Don't Know
Figure 3 for Constraining Representations Yields Models That Know What They Don't Know
Figure 4 for Constraining Representations Yields Models That Know What They Don't Know

Share this with someone who'll enjoy it:

A well-known failure mode of neural networks corresponds to high confidence erroneous predictions, especially for data that somehow differs from the training distribution. Such an unsafe behaviour limits their applicability. To counter that, we show that models offering accurate confidence levels can be defined via adding constraints in their internal representations. That is, we encode class labels as fixed unique binary vectors, or class codes, and use those to enforce class-dependent activation patterns throughout the model. Resulting predictors are dubbed Total Activation Classifiers (TAC), and TAC is used as an additional component to a base classifier to indicate how reliable a prediction is. Given a data instance, TAC slices intermediate representations into disjoint sets and reduces such slices into scalars, yielding activation profiles. During training, activation profiles are pushed towards the code assigned to a given training instance. At testing time, one can predict the class corresponding to the code that best matches the activation profile of an example. Empirically, we observe that the resemblance between activation patterns and their corresponding codes results in an inexpensive unsupervised approach for inducing discriminative confidence scores. Namely, we show that TAC is at least as good as state-of-the-art confidence scores extracted from existing models, while strictly improving the model's value on the rejection setting. TAC was also observed to work well on multiple types of architectures and data modalities.

Share this with someone who'll enjoy it: