CTL++: Evaluating Generalization on Never-Seen Compositional Patterns of Known Functions, and Compatibility of Neural Representations

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 9758 - 9767 December 7-11, 2022 ©2022 Association for Computational Linguistics CTL++: Evaluating Generalization on Never-Seen Compositional Patterns of Known Functions, and Compatibility of Neural Representations Róbert Csordás1 Kazuki Irie1 Jürgen Schmidhuber1,2 1The Swiss AI Lab IDSIA, USI & SUPSI, Lugano, Switzerland 2AI Initiative, KAUST, Thuwal, Saudi Arabia {robert, kazuki, juergen}@idsia.ch Abstract Well-designed diagnostic tasks have played a key role in studying the failure of neural nets (NNs) to generalize systematically. Famous ex- amples include SCAN and Compositional Ta- ble Lookup (CTL). Here we introduce CTL++, a new diagnostic dataset based on compositions of unary symbolic functions. While the orig- inal CTL is used to test length generalization or productivity, CTL++ is designed to test sys- tematicity of NNs, that is, their capability to generalize to unseen compositions of known functions. CTL++ splits functions into groups and tests performance on group elements com- posed in a way not seen during training. We show that recent CTL-solving Transformer vari- ants fail on CTL++. The simplicity of the task design allows for fine-grained control of task difficulty, as well as many insightful analyses. For example, we measure how much overlap between groups is needed by tested NNs for learning to compose. We also visualize how learned symbol representations in outputs of functions from different groups are compatible in case of success but not in case of failure. These results provide insights into failure cases reported on more complex compositions in the natural language domain. Our code is public.1 1 Introduction Neural networks (NNs) should ideally learn from training data to generalize systematically (Fodor et al., 1988), by learning generally applicable rules instead of pure pattern matching. Existing NNs, however, typically don’t. For example, in the con- text of sequence processing NNs, superficial differ- ences between training and test distributions, e.g., with respect to input sequence length or unseen input/word combinations, are enough to prevent current NNs from generalizing (Lake and Baroni, 2018). Training on a large amounts of data might alleviate the problem, but it is infeasible to cover all possible lengths and combinations. 1https://github.com/robertcsordas/ctlpp Indeed, while large language models trained on a large amounts of data have obtained impressive results (Brown et al., 2020), they often fail on tasks requiring simple algorithmic reasoning, e.g., simple arithmetics (Rae et al., 2021). A promis- ing way to achieve systematic generalization is to make NNs more compositional (Schmidhuber, 1990), by reflecting and exploiting the hierarchi- cal structure of many problems either within some NN’s learned weights, or through tailored NN ar- chitectures. For example, recent work by Csordás et al. (2022) proposes architectural modifications to the standard Transformer (Vaswani et al., 2017) motivated by the principles of compositionality. The resulting Neural Data Router (NDR) exhibits strong length generalization or productivity on rep- resentative datasets such as Compositional Table Lookup (CTL; Liska et al. (2018); Hupkes et al. (2019)). The focus of the present work is on systematicity: the capability to generalize …

Content truncated to 500 words