Adaptive token sampling for efficient transformer

Abstract

A transformer is described herein for using transformer-based technology to process data items (e.g., image items). The transformer increases the efficiency of the transformer-based technology by using a modified attention component. In operation, the modified attention component accepts embedding vectors that represent a plurality of item tokens, together with a classification token. A first stage of the modified attention component generates original attention information based on the embedding vectors. A second stage generates score information based on a portion of the original attention information that pertains to the classification token. A third stage produces modified attention information by removing attention values from the original attention information, as guided by a sampling operation that is performed on the score information. The second and third stages do not rely on machine-trained values, which expedites the deployment of these functions in existing transformers.