Radha Manisha Kopparti

Radha Kopparti, Tillman Weyde

Deep neural networks have become the dominant approach in natural language processing (NLP). However, in recent years, it has become apparent that there are shortcomings in systematicity that limit the performance and data efficiency of deep learning in NLP. These shortcomings can be clearly shown in lower-level artificial tasks, mostly on synthetic data. Abstract patterns are the best known examples of a hard problem for neural networks in terms of generalisation to unseen data. They are defined by relations between items, such as equality, rather than their values. It has been argued that these low-level problems demonstrate the inability of neural networks to learn systematically. In this study, we propose Embedded Relation Based Patterns (ERBP) as a novel way to create a relational inductive bias that encourages learning equality and distance-based relations for abstract patterns. ERBP is based on Relation Based Patterns (RBP), but modelled as a Bayesian prior on network weights and implemented as a regularisation term in otherwise standard network learning. ERBP is is easy to integrate into standard neural networks and does not affect their learning capacity. In our experiments, ERBP priors lead to almost perfect generalisation when learning abstract patterns from synthetic noise-free sequences. ERBP also improves natural language models on the word and character level and pitch prediction in melodies with RNN, GRU and LSTM networks. We also find improvements in in the more complex tasks of learning of graph edit distance and compositional sentence entailment. ERBP consistently improves over RBP and over standard networks, showing that it enables abstract pattern learning which contributes to performance in natural language tasks.

Radha Kopparti, Tillman Weyde

Learning abstract and systematic relations has been an open issue in neural network learning for over 30 years. It has been shown recently that neural networks do not learn relations based on identity and are unable to generalize well to unseen data. The Relation Based Pattern (RBP) approach has been proposed as a solution for this problem. In this work, we extend RBP by realizing it as a Bayesian prior on network weights to model the identity relations. This weight prior leads to a modified regularization term in otherwise standard network learning. In our experiments, we show that the Bayesian weight priors lead to perfect generalization when learning identity based relations and do not impede general neural network learning. We believe that the approach of creating an inductive bias with weight priors can be extended easily to other forms of relations and will be beneficial for many other learning tasks.

Radha Kopparti, Tillman Weyde

Many researchers implicitly assume that neural networks learn relations and generalise them to new unseen data. It has been shown recently, however, that the generalisation of feed-forward networks fails for identity relations.The proposed solution for this problem is to create an inductive bias with Differential Rectifier (DR) units. In this work we explore various factors in the neural network architecture and learning process whether they make a difference to the generalisation on equality detection of Neural Networks without and and with DR units in early and mid fusion architectures. We find in experiments with synthetic data effects of the number of hidden layers, the activation function and the data representation. The training set size in relation to the total possible set of vectors also makes a difference. However, the accuracy never exceeds 61% without DR units at 50% chance level. DR units improve generalisation in all tasks and lead to almost perfect test accuracy in the Mid Fusion setting. Thus, DR units seem to be a promising approach for creating generalisation abilities that standard networks lack.

Tillman Weyde, Radha Manisha Kopparti

In this paper, we show that standard feed-forward and recurrent neural networks fail to learn abstract patterns based on identity rules. We propose Relation Based Pattern (RBP) extensions to neural network structures that solve this problem and answer, as well as raise, questions about integrating structures for inductive bias into neural networks. Examples of abstract patterns are the sequence patterns ABA and ABB where A or B can be any object. These were introduced by Marcus et al (1999) who also found that 7 month old infants recognise these patterns in sequences that use an unfamiliar vocabulary while simple recurrent neural networks do not.This result has been contested in the literature but it is confirmed by our experiments. We also show that the inability to generalise extends to different, previously untested, settings. We propose a new approach to modify standard neural network architectures, called Relation Based Patterns (RBP) with different variants for classification and prediction. Our experiments show that neural networks with the appropriate RBP structure achieve perfect classification and prediction performance on synthetic data, including mixed concrete and abstract patterns. RBP also improves neural network performance in experiments with real-world sequence prediction tasks. We discuss these finding in terms of challenges for neural network models and identify consequences from this result in terms of developing inductive biases for neural network learning.

Tillman Weyde, Radha Manisha Kopparti

Basic binary relations such as equality and inequality are fundamental to relational data structures. Neural networks should learn such relations and generalise to new unseen data. We show in this study, however, that this generalisation fails with standard feed-forward networks on binary vectors. Even when trained with maximal training data, standard networks do not reliably detect equality.We introduce differential rectifier (DR) units that we add to the network in different configurations. The DR units create an inductive bias in the networks, so that they do learn to generalise, even from small numbers of examples and we have not found any negative effect of their inclusion in the network. Given the fundamental nature of these relations, we hypothesize that feed-forward neural network learning benefits from inductive bias in other relations as well. Consequently, the further development of suitable inductive biases will be beneficial to many tasks in relational learning with neural networks.