Memorize or generalize? Searching for a compositional RNN in a haystack

By Adam Liska, Germán Kruszewski, Marco Baroni. 2018

82 citations recorded by the source corpus.

Neural networks are very powerful learning systems, but they do not readily generalize from one task to the other. This is partly due to the fact that they do not learn in a compositional way, that is, by discovering skills that are shared by different tasks, and recombining them to solve new problems. In this paper, we explore the compositional generalization capabilities of recurrent neural networks (RNNs). We first propose the lookup table composition domain as a simple setup to test compositional behaviour and show that it is theoretically possible for a standard RNN to learn to behave compositionally in this domain when trained with standard gradient descent and provided with additional supervision. We then remove this additional supervision and perform a search over a large number of model initializations to investigate the proportion of RNNs that can still converge to a compositional solution. We discover that a small but non-negligible proportion of RNNs do reach partial compositional solutions even without special architectural constraints. This suggests that a combination of gradient descent and evolutionary strategies directly favouring the minority models that developed more compositional approaches might suffice to lead standard RNNs towards compositional solutions.

Subfields: Artificial Intelligence, Natural Language Processing, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Natural Language Processing

In Representation & Self-Supervised Learning

Citations in the corpus

Compositionality Decomposed: How do Neural Networks Generalise? (2019, background): “…the vast number of different tests that exist for compositionality [ Lake and Baroni, 2018; Hupkes et al. , 2018; Johnson et al. , 2017; Bahdanau et al. , 2018; Saxton et al. , 2019; Loula et al. , 2018; Dess`ı and Baroni, 2019; Liˇska et al. , 2018; Bowman et al. , 2015; Mul and Zuidema, 2019 ] .”
BERTs of a feather do not generalize together: Large variability in generalization across models with similar test set performance (2019, background): “Finally, Liˇska et al. (2018) trained 5000 instances of recurrent neural networks on the synthetic lookup tables task and found that most failed on compositional generalization but that a small number did display strong compositional generalization.”
Learning compositionally through attentive guidance (2018, methodology): “Training signal Following Weber, Shekhar, and Balasubramanian, we define the training signal on the target output that occurs in the dataset for a particular input sequence.”
Location Attention for Extrapolation to Longer Sequences (2019, methodology): “This issue has recently been uncovered for seq2seq models by looking at simple artiﬁcial tasks (Lake and Baroni, 2018; Liska et al., 2018; Weber et al., 2018).”
The Fine Line between Linguistic Generalization and Failure in Seq2Seq-Attention Models (2018, background): “Though the exact value of the initialization likely plays a factor, recent work by Liska et al. (2018) gives evidence that the order training data is presented may play an equally large part in determining whether a model achieves generalization.”
Transformers in Time-Series Analysis: A Tutorial (2022, background): “In some cases, the autoregressive nature of the training of machine learning models leads to memorization of past observations rather than the generalization of the training examples to new data [42, 52].”
The effect of sample size on different machine learning models for groundwater potential mapping in mountain bedrock aquifers (2020, background): “Adaptive Neuro-Fuzzy Inference System (ANFIS) 192 Artificial neural networks are powerful memorizing machines but are not capable of generalizing and 193 predicting patterns (Liška et al., 2018).”