On a Built-in Conflict between Deep Learning and Systematic Generalization
This study addresses a fundamental limitation in deep learning for researchers, potentially guiding new approaches to improve generalization, though it is incremental as it builds on existing concerns without proposing a solution.
The paper hypothesizes that internal function sharing in deep learning models weakens out-of-distribution and systematic generalization for classification tasks, showing this conflict occurs in various standard architectures like CNNs, ResNets, LSTMs, and Transformers.
In this paper, we hypothesize that internal function sharing is one of the reasons to weaken o.o.d. or systematic generalization in deep learning for classification tasks. Under equivalent prediction, a model partitions an input space into multiple parts separated by boundaries. The function sharing prefers to reuse boundaries, leading to fewer parts for new outputs, which conflicts with systematic generalization. We show such phenomena in standard deep learning models, such as fully connected, convolutional, residual networks, LSTMs, and (Vision) Transformers. We hope this study provides novel insights into systematic generalization and forms a basis for new research directions.