Such investigations have demonstrated that attention layers are able to implement a wide range of different algorithms, even for the same task, using both positional and semantic attributes of the inputs.

Such investigations have demonstrated that attention layers are able to implement a wide range of different algorithms, even for the same task, using both positional and semantic attributes of the inputs.

While machine learning (ML) models are increasingly used due to their high predictive power, their use in understanding the data-generating process (DGP) is limited. Understanding the DGP requires insights into feature-target associations, which many ML models cannot directly provide, due to their opaque internal mechanisms. Feature importance (FI) methods provide useful insights into the DGP under certain conditions. Since the results of different FI methods have different interpretations, selecting the correct FI method for a concrete use case is crucial and still requires expert knowledge. This paper serves as a comprehensive guide to help understand the different interpretations of FI methods. Through an extensive review of FI methods and providing new proofs regarding their interpretation, we facilitate a thorough understanding of these methods and formulate concrete recommendations for scientific inference. We conclude by discussing options for FI uncertainty estimation and point to directions for future research aiming at full statistical inference from black-box ML models.

We investigate how a dot-product attention layer learns a positional attention matrix (with tokens attending to each other based on their respective positions) and a semantic attention matrix (with tokens attending to each other based on their meaning). For an algorithmic task, we experimentally show how the same simple architecture can learn to implement a solution using either the positional or semantic mechanism. On the theoretical side, we study the learning of a non-linear self-attention layer with trainable tied and low-rank query and key matrices. In the asymptotic limit of high-dimensional data and a comparably large number of training samples, we provide a closed-form characterization of the global minimum of the non-convex empirical loss landscape. We show that this minimum corresponds to either a positional or a semantic mechanism and evidence an emergent phase transition from the former to the latter with increasing sample complexity. Finally, we compare the dot-product attention layer to linear positional baseline, and show that it outperforms the latter using the semantic mechanism provided it has access to sufficient data.