Finding Visual Task Vectors

Visual Prompting is a technique for teaching models to perform a visual taskvia in-context examples, without any additional training. In this work, weanalyze the activations of MAE-VQGAN, a recent Visual Prompting model, and findtask vectors, activations that encode task-specific information. Equipped withthis insight, we demonstrate that it is possible to identify the task vectorsand use them to guide the network towards performing different tasks withoutproviding any input-output examples. To find task vectors, we compute theaverage intermediate activations per task and use the REINFORCE algorithm tosearch for the subset of task vectors. The resulting task vectors guide themodel towards performing a task better than the original model without the needfor input-output examples.