# CNNs found to jump around more skillfully than RNNs: Compositional generalization in seq2seq convolutional networks

## TL;DR

This paper evaluates convolutional neural networks on compositional generalization tasks from the SCAN dataset, showing they outperform RNNs but still lack systematic rule induction, highlighting ongoing challenges in compositionality in neural models.

## Contribution

It demonstrates that CNNs can better handle compositional generalization than RNNs on seq2seq tasks, though they do not fully induce systematic rules.

## Key findings

- CNNs outperform RNNs on SCAN compositional tasks
- CNNs do not induce systematic compositional rules
- Performance gap highlights ongoing challenges in neural generalization

## Abstract

Lake and Baroni (2018) introduced the SCAN dataset probing the ability of seq2seq models to capture compositional generalizations, such as inferring the meaning of "jump around" 0-shot from the component words. Recurrent networks (RNNs) were found to completely fail the most challenging generalization cases. We test here a convolutional network (CNN) on these tasks, reporting hugely improved performance with respect to RNNs. Despite the big improvement, the CNN has however not induced systematic rules, suggesting that the difference between compositional and non-compositional behaviour is not clear-cut.