Paper ID: 2111.10361

Solving Visual Analogies Using Neural Algorithmic Reasoning

Atharv Sonwane, Gautam Shroff, Lovekesh Vig, Ashwin Srinivasan, Tirtharaj Dash

We consider a class of visual analogical reasoning problems that involve discovering the sequence of transformations by which pairs of input/output images are related, so as to analogously transform future inputs. This program synthesis task can be easily solved via symbolic search. Using a variation of the `neural analogical reasoning' approach of (Velickovic and Blundell 2021), we instead search for a sequence of elementary neural network transformations that manipulate distributed representations derived from a symbolic space, to which input images are directly encoded. We evaluate the extent to which our `neural reasoning' approach generalizes for images with unseen shapes and positions.

Submitted: Nov 19, 2021