The Devil is in the Detail: Simple Tricks Improve Systematic Generalization of Transformers
This work addresses the challenge of developing neural networks that generalize systematically, which is crucial for robust AI applications, though it is incremental as it builds on existing methods.
The paper tackles the problem of poor systematic generalization in Transformers by revisiting basic model configurations like embedding scaling and early stopping, resulting in accuracy improvements from 50% to 85% on PCFG and from 35% to 81% on COGS.
Recently, many datasets have been proposed to test the systematic generalization ability of neural networks. The companion baseline Transformers, typically trained with default hyper-parameters from standard tasks, are shown to fail dramatically. Here we demonstrate that by revisiting model configurations as basic as scaling of embeddings, early stopping, relative positional embedding, and Universal Transformer variants, we can drastically improve the performance of Transformers on systematic generalization. We report improvements on five popular datasets: SCAN, CFQ, PCFG, COGS, and Mathematics dataset. Our models improve accuracy from 50% to 85% on the PCFG productivity split, and from 35% to 81% on COGS. On SCAN, relative positional embedding largely mitigates the EOS decision problem (Newman et al., 2020), yielding 100% accuracy on the length split with a cutoff at 26. Importantly, performance differences between these models are typically invisible on the IID data split. This calls for proper generalization validation sets for developing neural networks that generalize systematically. We publicly release the code to reproduce our results.